Telefon-Nummer: +41 26 429 65 96
Description du projet :
Forschungsteam innerhalb von HES-SO:
Url des Projektstandortes:
Rey Joan Frédéric, Cesari Matias, Schoenenweid Marion, Montet Frédéric, Mauro Gandolla, Leevan Bonvin, Bourquin Vincent, Claude-Alain Jaccot, Justine Roman, Sebastian Duque Mahecha, Sergi Aguacil Moreno, Hennebert Jean, Goyette Pernot Joëlle
Journal of Physics: Conference Series, 2024 , vol.
Link zur Publikation
Radon is a noble, natural, and radioactive gas coming mainly from the ground which might accumulate indoors and lead each year to 200-300 deaths from lung cancer in Switzerland. A brand new and innovative living lab will be built as of 2023 in Fribourg (Switzerland) which will allow to tackle the built environment and the relationship with its occupants. Among a large panel of environmental parameters, radon gas will be continuously monitored under and around the building as well as in the building envelope. This paper aims to present the overall process of the radon dataflow: 1) design of the sensor probes, 2) implementation of the radon sensor probes in the ground and 3) go-live with the data sharing platform with the building users. Such an infrastructure will bring the opportunity to researchers to lead new and innovative radon-related research.
Frédéric Montet, Alessandro Pongelli, Joanathan Rial, Stefanie Schwab, Jean Hennebert, Thomas Jusselme
Acta Polytechnica CTU Proceedings,
2022, vol. 38: Central Europe towards sustainable building 2022 (CESB22), pp. 90-96
Building operation is responsible for 28% of the world’s carbon emissions. In this context, establishing priorities in refurbishment strategies at the scale of a city or a group of buildings is important. Such procedures are usually led by experts in energy performance and, therefore, they are rarely carried out due to their long and costly nature. This research aims at the estimation of building energy performance to pave the way towards finding near-optimal refurbishment strategies. Thanks to the identification of easily-accessible building characteristics, the method applies machine learning models to scan a building portfolio based on a low level of details. The results show good potential to identify low-performer buildings with simple machine learning methods. It also opens the door for further improvements through the inclusion of supplementary building features at the input of the predictive system. This work includes (a) the integration of a knowledge database thanks to the Swiss CECB energy performance certificates, referencing more than 70 000 buildings, (b) the preparation of a training data set through the selection of relevant physical characteristics of buildings (input) and the corresponding energy consumption labels (output), (c) the development of predictive models used in a supervised way, (d) their evaluation on an independent test set.
Lucy Linder, Frédéric Montet, Jean Hennebert, Jean-Philippe Bacher
Journal of Physics: Conference Series,
2021, no. 2042, article no. 012016
The modern built environment is now connected. Multiple software and protocols are used in buildings of many kinds, thus creating a fascinating and heterogeneous environment. Within this context, applied research can be complicated and would benefit from a single data location across projects and users. The first version of BBData tried to solve this problem, BBData v2.0 is an update with a better-defined scope and a new codebase.
The solution has been open sourced and simplified with a full software rewrite. Its components are now state-of-the-art and proven to be stable in industrial settings. The achieved performances have been thoroughly tested. Together with its new architecture, BBData v2.0 now accommodates the needs of modern experiments; efficient for simple proof of concepts while keeping the possibility to scale up to city-level projects. This flexibility makes BBData a good candidate for research while being able to scale in production settings.
Frédéric Montet, Lorenz Rychener, Alessandro Pongelli, Jean Hennebert, Jean-Philippe Bacher
Journal of Physics: Conference Series,
2021, no. 2042, article no. 012026
With the fourth generation of district heating networks in sight, opportunities are rising for better services and optimized planning of energy production. Indeed, the more intensive data collection is expected to allow for load prediction, customer profiling, etc. In this context, our work aims at a better understanding of customer profiles from the captured data.
Given the variety of households, such profiles are difficult to capture. This study explores the possibility to predict domestic hot water (DHW) usage. Such prediction is made challenging due to the presence of two components in the signal, the first one bound to the physical properties of the DHW distribution system, the second one bound to the human patterns related to DHW consumption.
Our contributions include (1) the analysis of recurrent neural network architectures based on GRU, (2) the inclusion of state-based labels inferred in an unsupervised way to simulate domain knowledge, (3) the comparison of different features.
Results show that the physical contribution in the signal can be forecasted successfully across households. On the contrary, the stochastic "human" component is harder to predict and would need further research, either by improving the modelling or by including alternate signals.
Jonathan Parrat, Jean-Philippe Bacher, Florinel Radu, Jean Hennebert
bulletin.ch = Fachzeitschrift und Verbandsinformationen von Electrosuisse und VSE = Bulletin SEV/AES : revue spécialisée et informations des associations Electrosuisse et AES,
2020, vol. 6, pp. 22-26
Lorenz Rychener, Frédéric Montet, Jean Hennebert
Procedia Computer Science,
2020, vol. 170, pp. 648-655
In the context of Industry 4.0, an emerging trend is to increase the reliability of industrial process by using machine learning (ML) to detect anomalies of production machines. The main advantages of ML are in the ability to (1) capture non-linear phenomena, (2) adapt to many different processes without human intervention and (3) learn incrementally and improve over time. In this paper, we take the perspective of IT system architects and analyse the implications of the inclusion of ML components into a traditional anomaly detection systems. Through a prototype that we deployed for chemical reactors, our findings are that such ML components are impacting drastically the architecture of classical alarm systems. First, there is a need for long-term storage of the data that are used to train the models. Second, the training and usage of ML models can be CPU intensive and may request using specific resources. Third, there is no single algorithm that can detect machine errors. Fourth, human crafted alarm rules can now also include a learning process to improve these rules, for example by using active learning with a human-in-the-loop approach. These reasons are the motivations behind a microservice-based architecture for an alarm system in industrial machinery.
Wolf Beat, Donzallaz Jonathan, Buchs Colette, Hayoz Amanda, Commend Stéphane, Hennebert Jean
Proceedings of IAPR Workshop on Artificial Neural Networks in Pattern Recognition, ANNPR 2020 : Artificial Neural Networks in Pattern Recognition, 2-4 September 2020, Winterthur, Switzerland, 2020
Deep excavations are today mainly designed by manually optimising the wall’s geometry, stiffness and strut or anchor layout. In order to better automate this process for sustained excavations, we are exploring the possibility of approximating key values using a machine learning (ML) model instead of calculating them with time-consuming numerical simulations. After demonstrating in our previous work that this approach works for simple use cases, we show in this paper that this method can be enhanced to adapt to complex real-world supported excavations. We have improved our ML model compared to our previous work by using a convolutional neural network (CNN) model, coding the excavation configuration as a set of layers of fixed height containing the soil parameters as well as the geometry of the walls and struts. The system is trained and evaluated on a set of synthetically generated situations using numerical simulation software. To validate this approach, we also compare our results to a set of 15 real-world situations in a t-SNE. Using our improved CNN model we could show that applying machine learning to predict the output of numerical simulation in the domain of geotechnical engineering not only works in simple cases but also in more complex, real-world situations.
Rychener Lorenz, Montet Frédéric, Hennebert Jean
Procedia Computer Science, 2020 , vol.
In this paper, we take the perspective of IT system architects and analyse the implications of the inclusion of ML components into a traditional anomaly detection systems. Through a prototype that we deployed for chemical reactors, our findings are that such ML components are impacting drastically the architecture of classical alarm systems.
Commend Stéphane, Sacha Wattel, Hennebert Jean, Kuonen Pierre, Laurent Vulliet
Proceedings of XIV International Conference on Computational Plasticity, Fundamentals and Application (COMPLAS 2019), 2019 , pp.
Keith R. Martin, Kaweh Mansouri, Robert N. Weinreb, Robert Wasilewicz, Christophe Gisler, Jean Hennebert, Dominique Genoud
American Journal of Ophthalmology,
2018, vol. 194, pp. 46-53
Purpose : To test the hypothesis that contact lens sensor (CLS)-based 24-hour profiles of ocular volume changes contain information complementary to intraocular pressure (IOP) to discriminate between primary open-angle glaucoma (POAG) and healthy (H) eyes.
Design : Development and evaluation of a diagnostic test with machine learning.
Methods : Subjects: From 435 subjects (193 healthy and 242 POAG), 136 POAG and 136 age-matched healthy subjects were selected. Subjects with contraindications for CLS wear were excluded. Procedure: This is a pooled analysis of data from 24 prospective clinical studies and a registry. All subjects underwent 24-hour CLS recording on 1 eye. Statistical and physiological CLS parameters were derived from the signal recorded. CLS parameters frequently associated with the presence of POAG were identified using a random forest modeling approach. Main Outcome Measures: Area under the receiver operating characteristic curve (ROC AUC) for feature sets including CLS parameters and Start IOP, as well as a feature set with CLS parameters and Start IOP combined.
Results : The CLS parameters feature set discriminated POAG from H eyes with mean ROC AUCs of 0.611, confidence interval (CI) 0.493–0.722. Larger values of a given CLS parameter were in general associated with a diagnosis of POAG. The Start IOP feature set discriminated between POAG and H eyes with a mean ROC AUC of 0.681, CI 0.603–0.765. The combined feature set was the best indicator of POAG with an ROC AUC of 0.759, CI 0.654–0.855. This ROC AUC was statistically higher than for CLS parameters or Start IOP feature sets alone (both P < .0001).
Conclusions : CLS recordings contain information complementary to IOP that enable discrimination between H and POAG. The feature set combining CLS parameters and Start IOP provide a better indication of the presence of POAG than each of the feature sets separately. As such, the CLS may be a new biomarker for POAG.
Oussama Zayene, Sameh Masmoudi Touj, Jean Hennebert, Rolf Ingold, Najoua Essoukri Ben Amara
IET Computer Vision,
2018, vol. 12, no. 5, pp. 710-719
This study presents a novel approach for Arabic video text recognition based on recurrent neural networks. In fact, embedded texts in videos represent a rich source of information for indexing and automatically annotating multimedia documents. However, video text recognition is a non-trivial task due to many challenges like the variability of text patterns and the complexity of backgrounds. In the case of Arabic, the presence of diacritic marks, the cursive nature of the script and the non-uniform intra/inter word distances, may introduce many additional challenges. The proposed system presents a segmentation-free method that relies specifically on a multi-dimensional long short-term memory coupled with a connectionist temporal classification layer. It is shown that using an efficient pre-processing step and a compact representation of Arabic character models brings robust performance and yields a low-error rate than other recently published methods. The authors’ system is trained and evaluated using the public AcTiV-R dataset under different evaluation protocols. The obtained results are very interesting. They also outperform current state-of-the-art approaches on the public dataset ALIF in terms of recognition rates at both character and line levels.
Baptiste Wicht, Andreas Fischer, Jean Hennebert
Dans Fischer, Andreas, Hennebert, Jean, Wicht, Baptiste, Lecture Notes in Computer Science
(12 p.). 2018,
Cham : Springer
Link zur Publikation
Deep Learning Library (DLL) is a library for machine learning with deep neural networks that focuses on speed. It supports feedforward neural networks such as fully-connected Artificial Neural Networks (ANNs) and Convolutional Neural Networks (CNNs). Our main motivation for this work was to propose and evaluate novel software engineering strategies with potential to accelerate runtime for training and inference. Such strategies are mostly independent of the underlying deep learning algorithms. On three different datasets and for four different neural network models, we compared DLL to five popular deep learning libraries. Experimentally, it is shown that the proposed library is systematically and significantly faster on CPU and GPU. In terms of classification performance, similar accuracies as the other libraries are reported.
Journal of Imaging,
2018, vol. 4(2), no. 32
Recognizing texts in video is more complex than in other environments such as scanned documents. Video texts appear in various colors, unknown fonts and sizes, often affected by compression artifacts and low quality. In contrast to Latin texts, there are no publicly available datasets which cover all aspects of the Arabic Video OCR domain. This paper describes a new well-defined and annotated Arabic-Text-in-Video dataset called AcTiV 2.0. The dataset is dedicated especially to building and evaluating Arabic video text detection and recognition systems. AcTiV 2.0 contains 189 video clips serving as a raw material for creating 4063 key frames for the detection task and 10,415 cropped text images for the recognition task. AcTiV 2.0 is also distributed with its annotation and evaluation tools that are made open-source for standardization and validation purposes. This paper also reports on the evaluation of several systems tested under the proposed detection and recognition protocols.
Keith R. Martin, Kaweh Mansouri, Robert N. Weinreb, Robert Wasilevicz, Gisler Christophe, Hennebert Jean, Genoud Dominique
American journal of ophthalmology, 2018 , vol.
Florian Rossier, Philippe Lang, Jean Hennebert
2017, vol. 122, pp. 691-696
Electricity load monitoring in residential buildings has become an important task allowing for energy consumption understanding, indirect human activity recognition and occupancy modelling. In this context, Non Intrusive Load Monitoring (NILM) is an approach based on the analysis of the global electricity consumption signal of the habitation. Current NILM solutions are reaching good precision for the identification of electrical devices but at the cost of difficult setups with expensive equipments typically working at high frequency. In this work we propose to use a low-cost and easy to install low frequency sensor for which we improve the performances with an active machine learning strategy. At setup, the system is able to identify some appliances with typical signatures such as a fridge. During usage, the system detects unknown signatures and provides a user-friendly procedure to include new appliances and to improve the identification precision over time.
Lucy Linder, Damien Vionnet, Jean-Philippe Bacher, Jean Hennebert
2017, vol. 122, pp. 589-594
Future buildings will more and more rely on advanced Building Management Systems (BMS) connected to a variety of sensors, actuators and dedicated networks. Their objectives are to observe the state of rooms and apply automated rules to preserve or increase comfort while economizing energy. In this work, we advocate for the inclusion of a dedicated system for sensors data storage and processing, based on Big Data technologies. This choice enables new potentials in terms of data analytics and applications development, the most obvious one being the ability to scale up seamlessly from one smart building to several, in the direction of smart areas and smart cities. We report in this paper on our system architecture and on several challenges we met in its elaboration, attempting to meet requirements of scalability, data processing, flexibility, interoperability and privacy. We also describe current and future end-user services that our platform will support, including historical data retrieval, visualisation, processing and alarms. The platform, called BBData - Big Building Data, is currently in production at the Smart Living Lab of Fribourg and is offered to several research teams to ease their work, to foster the sharing of historical data and to avoid that each project develops its own data gathering and processing pipeline.
Antonio Ridi, Gisler Christophe, Hennebert Jean
2016 23rd International Conference on Pattern Recognition (ICPR), 2016
Damien Zufferey, Thomas Hofer, Jean Hennebert, Michael Schumacher, Rolf Ingold, Stefano Bromuri
Computers in biology and medicine,
October 2015, vol. 65, pp. 34–43
We are motivated by the issue of classifying diseases of chronically ill patients to assist physicians in their everyday work. Our goal is to provide a performance comparison of state-of-the-art multi-label learning algorithms for the analysis of multivariate sequential clinical data from medical records of patients affected by chronic diseases. As a matter of fact, the multi-label learning approach appears to be a good candidate for modeling overlapped medical conditions, specific to chronically ill patients. With the availability of such comparison study, the evaluation of new algorithms should be enhanced. According to the method, we choose a summary statistics approach for the processing of the sequential clinical data, so that the extracted features maintain an interpretable link to their corresponding medical records. The publicly available MIMIC-II dataset, which contains more than 19,000 patients with chronic diseases, is used in this study. For the comparison we selected the following multi-label algorithms: ML-kNN, AdaBoostMH, binary relevance, classifier chains, HOMER and RAkEL. Regarding the results, binary relevance approaches, despite their elementary design and their independence assumption concerning the chronic illnesses, perform optimally in most scenarios, in particular for the detection of relevant diseases. In addition, binary relevance approaches scale up to large dataset and are easy to learn. However, the RAkEL algorithm, despite its scalability problems when it is confronted to large dataset, performs well in the scenario which consists of the ranking of the labels according to the dominant disease of the patient.
Kai Chen, Mathias Seuret, Hao Wei, Marcus Liwicki, Jean Hennebert, Rolf Ingold
Proceedings of Document Recognition and Retrieval XXII; 940204 (2015), 8-12 February 2015, San Francisco, California, USA,
2015, vol. 9402
In this paper, we propose a new dataset and a ground-truthing methodology for layout analysis of historical documents with complex layouts. The dataset is based on a generic model for ground-truth presentation of the complex layout structure of historical documents. For the purpose of extracting uniformly the document contents, our model defines five types of regions of interest: page, text block, text line, decoration, and comment. Unconstrained polygons are used to outline the regions. A performance metric is proposed in order to evaluate various page segmentation methods based on this model. We have analysed four state-of-the-art ground-truthing tools: TRUVIZ, GEDI, WebGT, and Aletheia. From this analysis, we conceptualized and developed Divadia, a new tool that overcomes some of the drawbacks of these tools, targeting the simplicity and the efficiency of the layout ground truthing process on historical document images. With Divadia, we have created a new public dataset. This dataset contains 120 pages from three historical document image collections of different styles and is made freely available to the scientific community for historical document layout analysis research.
Antonio Ridi, Zarkadis Nikos, Gisler Christophe, Hennebert Jean
2015 IEEE International Conference on Data Science and Advanced Analytics (DSAA), 2015 , pp.
2015 IEEE wireless communications and networking conference workshops (WCNCW), 2015 , pp.
Gisler Christophe, Antonio Ridi, Hennebert Jean, Robert N. Weinreb, Kaweh Mansouri
Translational vision science & technology, 2015 , vol.
Christophe Gisler, Antonio Ridi, Jean Hennebert, Robert N. Winreb, Kaweh Mansouri
Translational Vision Science Technology (TVST),
2015, vol. 4, no. 1, article no. 4
To detect and quantify eye blinks during 24-hour intraocular pressure (IOP) monitoring with a contact lens sensor (CLS). A total of 249 recordings of 24-hour IOP patterns from 202 participants using a CLS were included. Software was developed to automatically detect eye blinks, and wake and sleep periods. The blink detection method was based on detection of CLS signal peaks greater than a threshold proportional to the signal amplitude. Three methods for automated detection of the sleep and wake periods were evaluated. These relied on blink detection and subsequent comparison of the local signal amplitude with a threshold proportional to the mean signal amplitude. These methods were compared to manual sleep/wake verification. In a pilot, simultaneous video recording of 10 subjects was performed to compare the software to observer-measured blink rates. Mean (SD) age of participants was 57.4 ± 16.5 years (males, 49.5%). There was excellent agreement between software-detected number of blinks and visually measured blinks for both observers (intraclass correlation coefficient [ICC], 0.97 for observer 1; ICC, 0.98 for observer 2). The CLS measured a mean blink frequency of 29.8 ± 15.4 blinks/min, a blink duration of 0.26 ± 0.21 seconds and an interblink interval of 1.91 ± 2.03 seconds. The best method for identifying sleep periods had an accuracy of 95.2 ± 0.5%. Automated analysis of CLS 24-hour IOP recordings can accurately quantify eye blinks, and identify sleep and wake periods. This study sheds new light on the potential importance of eye blinks in glaucoma and may contribute to improved understanding of circadian IOP characteristics.
Michael Schumacher, Stefano Bromuri, Damien Zufferey, Jean Hennebert
Journal of biomedical informatics,
octobre 2014, vol. 51, pp. 165-175
Objective This research is motivated by the issue of classifying illnesses of chronically ill patients for decision support in clinical settings. Our main objective is to propose multi-label classification of multivariate time series contained in medical records of chronically ill patients, by means of quantization methods, such as bag of words (BoW), and multi-label classification algorithms. Our second objective is to compare supervised dimensionality reduction techniques to state-of-the-art multi-label classification algorithms. The hypothesis is that kernel methods and locality preserving projections make such algorithms good candidates to study multi-label medical time series. Methods We combine BoW and supervised dimensionality reduction algorithms to perform multi-label classification on health records of chronically ill patients. The considered algorithms are compared with state-of-the-art multi-label classifiers in two real world datasets. Portavita dataset contains 525 diabetes type 2 (DT2) patients, with co-morbidities of DT2 such as hypertension, dyslipidemia, and microvascular or macrovascular issues. MIMIC II dataset contains 2635 patients affected by thyroid disease, diabetes mellitus, lipoid metabolism disease, fluid electrolyte disease, hypertensive disease, thrombosis, hypotension, chronic obstructive pulmonary disease (COPD), liver disease and kidney disease. The algorithms are evaluated using multi-label evaluation metrics such as hamming loss, one error, coverage, ranking loss, and average precision. Results Non-linear dimensionality reduction approaches behave well on medical time series quantized using the BoW algorithm, with results comparable to state-of-the-art multi-label classification algorithms. Chaining the projected features has a positive impact on the performance of the algorithm with respect to pure binary relevance approaches. Conclusions The evaluation highlights the feasibility of representing medical health records using the BoW for multi-label classification tasks. The study also highlights that dimensionality reduction algorithms based on kernel methods, locality preserving projections or both are good candidates to deal with multi-label classification tasks in medical time series with many missing values and high label density.
Gisler Christophe, Antonio Ridi, Milène Fauquex, Genoud Dominique, Hennebert Jean
The 6th International Conference of Soft Computing and Pattern Recognition (SoCPaR 2014), 2014
2014 international conference on data science and advanced analytics (DSAA), 2014
22nd international conference on pattern recognition, 2014
Gérôme Bovet, Antonio Ridi, Jean Hennebert
Dans Bessis, Nik, Dobre, Ciprian, Big data and internet of things: a roadmap for smart environments
(pp. 259-283). 2014,
Cham : Springer
The emerging concept of Smart Building relies on an intensive use of sensors and actuators and therefore appears, at first glance, to be a domain of predilection for the IoT. However, technology providers of building automation systems have been functioning, for a long time, with dedicated networks, communication protocols and APIs. Eventually, a mix of different technologies can even be present in a given building. IoT principles are now appearing in buildings as a way to simplify and standardise application development. Nevertheless, many issues remain due to this heterogeneity between existing installations and native IP devices that induces complexity and maintenance efforts of building management systems. A key success factor for the IoT adoption in Smart Buildings is to provide a loosely-coupled Web protocol stack allowing interoperation between all devices present in a building. We review in this chapter different strategies that are going in this direction. More specifically, we emphasise on several aspects issued from pervasive and ubiquitous computing like service discovery. Finally, making the assumption of seamless access to sensor data through IoT paradigms, we provide an overview of some of the most exciting enabling applications that rely on intelligent data analysis and machine learning for energy saving in buildings.
Progress in Pattern Recognition, Image Analysis, Computer Vision, and Applications, 2013 , vol.
Gisler Christophe, Antonio Ridi, Hennebert Jean
The 8th International Workshop on Systems, Signal Processing and their Applications 2013 (WoSSPA 2013), 2013
Damien Zufferey, Gisler Christophe, Abou Khaled Omar, Hennebert Jean
The 11th International Conference on Information Science, Signal Processing and their Applications (ISSPA 2012), 2012 , pp.
Frédéric Montet, Alessandro Pongelli, Stefanie Schwab, Mylène Devaux, Thomas Jusselme, Jean Hennebert
Proceedings of cisbat 2023, the built environment in transititon, Hybrid International Scientific Conference, 13-15 September 2023, Lausanne, Switzerland
Link zur Konferenz
This paper presents an innovative methodology for enhancing energy efficiency assessment procedures in the built environment, with a focus on the Switzerland’s Energy Strategy 2050. The current methodology necessitates intensive expert surveys, leading to substantial time and cost implications. Also, such a process can’t be scaled to a large number of buildings. Using machine learning techniques, the estimation process is augmented and exploit open data resources. Utilizing a robust dataset exceeding 70’000 energy performance certificates (CECB), the method devises a two-stage ML approach to forecast energy performance. The first phase involves data reconstruction from online repositories, while the second employs a regression algorithm to estimate the energy efficiency. The proposed approach addresses the limitations of existing machine learning methods by offering finer prediction granularity and incorporating readily available data. The results show a commendable degree of prediction accuracy, particularly for single-family residences. Despite this, the study reveals a demand for further granular data, and underlines privacy concerns associated with such data collection. In summary, this investigation provides a significant contribution to the enhancement of energy efficiency assessment methodologies and policymaking.
Joan Frédéric Rey, Matias Cesari, Frédéric Montet, M. Gandolla, Leewan Bonvin, Vincent Bourquin, C. L. Jacot, Justine Roman, Sebastian Duque Mahecha, Sergi Aguacil Moreno, Jean Hennebert, Joëlle Goyette Pernot
Journal of Physics: Conference Series ; Proceedings of cisbat 2023, the built environment in transititon, Hybrid International Scientific Conference, 13-15 September 2023, Lausanne, Switzerland
Jean Luc Bloechle, Jean Hennebert, Christophe Gisler
Proceedings of the ACM Symposium on Document Engineering (DocEng'23), 22-25 August 2023, Limerick, Ireland
Optical Character Recognition (OCR) from document photos taken by cell phones is a challenging task. Most OCR methods require prior binarization of the image, which can be difficult to achieve when documents are captured with various mobile devices in unknown lighting conditions. For example, shadows cast by the camera or the camera holder on a hard copy can jeopardize the binarization process and hinder the next OCR step. In the case of highly uneven illumination, binarization methods using global thresholding simply fail, and state-of-the-art adaptive algorithms often deliver unsatisfactory results. In this paper, we present a new binarization algorithm using two complementary local adaptive passes and taking advantage of the color components to improve results over current image binarization methods. The proposed approach gave remarkable results at the DocEng'22 competition on the binarization of photographed documents.
Teo Brigljevic, Jean-Philippe Bacher, Jean Hennebert
Proceedings of the 6th European Grid Service Market Symposium, 4-5 July 2022, Lucerne, Switzerland
Among the issues distribution system operators are anticipating today, the management of peak loads is at the forefront. This problematic has repercussions on the dimensioning of the network infrastructure that must support these peaks, consequence of the deployment of distribution energy resources like solar panels, or the introduction of future consumption needs driven by the electrification of the energy system. In answer to this problematic, leveraging the electrical flexibility of end-customers mixing consumption and production profiles is considered particularly promising to avoid the oversizing of the grid infrastructure or the use of dimensioning cases too strict for connecting new end-customers to the grid. To determine the flexibility shares the DSOs can use as leeway for the grid management, the end-customers’ load profiles made available by the deployment of advance metering infrastructure are leveraged, in combination with the data obtained from the DSOs’ geographical and network information systems. Utilizing the deployment of their advance metering infrastructure, Groupe E initiated a data-driven project analyzing the load profiles in a typical low-voltage distribution grid covering 269 distinct end-customers. These end-customers present combinations of profiles mixing baseline residential consumption, particular non-standard consumption (heat pumps, electrical vehicles) and production (photovoltaic) profiles. Based on this data, a methodology has been proposed to identify critical loci in time and location where the grid infrastructure reaches its load limits and targets select impacted end-customers for the activation of flexibility lowering in return the peaks critical for the infrastructure. The methodology follows a data-oriented approach to (1) identify the flexibility shares in the grid where the activation improves the load on the infrastructure using the load profiles and (2) propose a flexibility controlling strategy benefiting the grid infrastructure with the analysis of the effects of this strategy on the grid infrastructure and the impacted end-customers.
Frédéric Montet, Alessandro Pongelli, Rial, Stefanie Schwab, Jean Hennebert, Thomas Jusselme
Proceedings of Central Europe towards Sustainable Building 2022, International Scientific Conference, 4-6 July 2022, Prague, Czech Republic; Acta Polytechnica
Building operation is responsible for 28% of the world’s carbon emissions. In this context, establishing priorities in refurbishment strategies at the scale of a city or a group of buildings is important. Such procedures are usually led by experts in energy performance and, therefore, they are rarely carried out due to their long and costly nature. This research aims at the estimation of building energy performance to pave the way towards finding near-optimal refurbishment strategies. Thanks to the identification of easily-accessible building characteristics, the method applies machine learning models to scan a building portfolio based on a low level of details. The results show good potential to identify low-performer buildings with simple machine learning methods. It also opens the door for further improvements through the inclusion of supplementary building features at the input of the predictive system. This work includes (a) the integration of a knowledge database thanks to the Swiss CECB energy performance certificates, referencing more than 70’000 buildings, (b) the preparation of a training data set through the selection of relevant physical characteristics of buildings (input) and the corresponding energy consumption labels (output), (c) the development of predictive models used in a supervised way, (d) their evaluation on an independent test set.
Oussama Zayene, Rolf Ingold, Najoua Essoukri BenAmara, Jean Hennebert
Proceedings of International Conference on Pattern Recognition, ICPR 2021: Pattern Recognition, CPR International Workshops and Challenges, 10-15 January 2021, Virtual Event
After the success of the two first editions of the “Arabic Text in Videos Competition—AcTiVComp”, we are proposing to organize a new edition in conjunction with the 25th International Conference on Pattern Recognition (ICPR’20). The main objective is to contribute in the research field of text detection and recognition in multimedia documents, with a focus on Arabic text in video frames. The former editions were held in the framework of ICPR’16 and ICDAR’17 conferences. The obtained results on the AcTiV dataset have shown that there is still room for improvement in both text detection and recognition tasks. Four groups with five systems are participating to this edition of AcTiVComp (three for the detection task and two for the recognition task). All the submitted systems have followed a CRNN-based architecture, which is now the de facto choice for text detection and OCR problems. The achieved results are very interesting, showing a significant improvement from the state-of-the-art performances on this field of research.
Zayene Oussama, Najoua Essoukri BenAmara, Rolf Ingold, Hennebert Jean
International Conference on Pattern Recognition, 10.01.2021 - 15.01.2021, Milan
After the success of the two first editions of the Arabic Text in Videos Competition|AcTiVComp" , we are proposing to organize a new edition in conjunction with the 25th International Conference on Pattern Recognition (ICPR'20). The main objective is to contribute in the research eld of text detection and recognition in multimedia documents, with a focus on Arabic text in video frames. The former editions were held in the framework of ICPR'16 and ICDAR'17 conferences. The obtained results on the AcTiV dataset have shown that there is still room for improvement in both text detection and recognition tasks. Four groups with ve systems are participating to this edition of AcTiV-Comp (three for the detection task and two for the recognition task). All the submitted systems have followed a CRNN-based architecture, which is now the de facto choice for text detection and OCR problems. The achieved results are very interesting, showing a signicant improvement from the state-of-the-art performances on this eld of research.
Beat Wolf, Jonathan Donzallaz, Colette Jost, Amanda Hayoz, Stéphane Commend, Jean Hennebert
Lecture Notes in Computer Science ; Proceedings of IAPR Workshop on Artificial Neural Networks in Pattern Recognition, ANNPR 2020 : Artificial Neural Networks in Pattern Recognition, 2-4 September 2020, Winterthur, Switzerland
Lucy Linder, Michael Jungo, Jean Hennebert, Claudiu Musat, Andreas Fischer
Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), 11-16 May 2020, Marseille, France
This paper presents SwissCrawl, the largest Swiss German text corpus to date. Composed of more than half a million sentences, it was generated using a customized web scraping tool that could be applied to other low-resource languages as well. The approach demonstrates how freely available web pages can be used to construct comprehensive text corpora, which are of fundamental importance for natural language processing. In an experimental evaluation, we show that using the new corpus leads to significant improvements for the task of language modeling. To capture new content, our approach will run continuously to keep increasing the corpus over time.
XIV International Conference on Computational Plasticity, Fundamentals and Application (COMPLAS 2019), 03.09.2019 - 05.09.2019, Barcelona
Proceedings of the 2018 International Conference on High Performance Computing & Simulation (HPCS 2018), The 16th Annual Meeting, 16-20 July 2018, Orléans, France
Expression Templates is a technique allowing to write linear algebra code in C++ the same way it would be written on paper. It is also used extensively as a performance optimization technique, especially as the Smart Expression Templates form which allows for even higher performance. It has proved to be very efficient for computation on a Central Processing Unit (CPU). However, due to its design, it is not easily implemented on a Graphics Processing Unit (GPU). In this paper, we devise a set of techniques to allow the seamless evaluation of Smart Expression Templates on the GPU. The execution is transparent for the user of the library which still uses the matrices and vector as if it was on the CPU and profits from the performance and higher multi-processing capabilities of the GPU. We also show that the GPU version is significantly faster than the CPU version, without any change to the code of the user.
Oussama Zayene, Jean Hennebert, Rolf Ingold, Najoua Essoukri Ben Amara
Proceedings of 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), 9-15 November 2017, Kyoto, Japan
This paper describes the multi-resolution Arabic Text detection and recognition in Video Competition-AcTiVComp held in the context of the 14 th International Conference on Document Analysis and Recognition (ICDAR' 2017), during November 9-15, 2017, in Kyoto, Japan. The main objective of this competition is to evaluate the performance of participants' algorithms for automatically detecting and recognizing Arabic texts in video frames using the freely available Arabic-Text-in-Video (AcTiV) dataset. A first edition was held in the framework of the 23 rd International Conference on Pattern Recognition (ICPR'2016). Three groups with five systems are participating to the second edition of AcTiVComp. These systems are tested in a blind manner on a closed-subset of the AcTiV database, which is unknown to all participants. In addition to the experimental setup and observed results, we also provide a short description of the participating groups and their systems.
Kai Chen, Mathias Seuret, Jean Hennebert, Rolf Ingold
Proceedings of 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), 9-15 November 2017, Kyoto, Japan
This paper presents a page segmentation method for handwritten historical document images based on a Convolutional Neural Network (CNN). We consider page segmentation as a pixel labeling problem, i.e., each pixel is classified as one of the predefined classes. Traditional methods in this area rely on hand-crafted features carefully tuned considering prior knowledge. In contrast, we propose to learn features from raw image pixels using a CNN. While many researchers focus on developing deep CNN architectures to solve different problems, we train a simple CNN with only one convolution layer. We show that the simple architecture achieves competitive results against other deep architectures on different public datasets. Experiments also demonstrate the effectiveness and superiority of the proposed method compared to previous methods.
Fouad Slimane, Rolf Ingold, Jean Hennebert
Abstract:This paper describes the organisation and results of the Arabic Recognition Competition: Multi-font Multi-size Digitally Represented Text held in the context of the 14th International Conference on Document Analysis and Recognition (ICDAR'2017), during November 10-15, 2017, Kyoto, Japan. This competition has used the freely available Arabic Printed Text Image (APTI) database. A first and second editions took place respectively in ICDAR'2011 and ICDAR'2013. In this edition, we propose four challenges. Six research groups are participating in the competition with thirteen systems. These systems are compared using the font, font-size, font and font-size, and character and word recognition rates. The systems were tested in a blind manner using the first 5000 images of APTI database set 6. A short description of the participating groups, their systems, the experimental setup, and the observed results are presented.
Antonio Ridi, Christophe Gisler, Jean Hennebert
Proceedings of 2016 23rd International Conference on Pattern Recognition (ICPR), 4-8 December 2016, Cancun, Mexico
In this work we provide details on a new and effective approach able to generate Gaussian Mixture Models (GMMs) for the classification of aggregated time series. More specifically, our procedure can be applied to time series that are aggregated together by adding their features. The procedure takes advantage of the additive property of the Gaussians that complies with the additive property of the features. Our goal is to classify aggregated time series, i.e. we aim to identify the classes of the single time series contributing to the total. The standard approach consists in training the models using the combination of several time series coming from different classes. However, this has the drawback of being a very slow operation given the amount of data. The proposed approach, called GMMs aggregation procedure, addresses this problem. It consists of three steps: (i) modeling the independent classes, (ii) generation of the models for the class combinations and (iii) simplification of the generated models. We show the effectiveness of our approach by using time series in the context of electrical appliance consumption, where the time series are aggregated by adding the active and reactive power. Finally, we compare the proposed approach with the standard procedure.
Proceedings of the 2016 23rd International Conference on Pattern Recognition (ICPR), 4-8 December 2016, Cancun, Mexico
Deep learning had a significant impact on diverse pattern recognition tasks in the recent past. In this paper, we investigate its potential for keyword spotting in handwritten documents by designing a novel feature extraction system based on Convolutional Deep Belief Networks. Sliding window features are learned from word images in an unsupervised manner. The proposed features are evaluated both for template-based word spotting with Dynamic Time Warping and for learning-based word spotting with Hidden Markov Models. In an experimental evaluation on three benchmark data sets with historical and modern handwriting, it is shown that the proposed learned features outperform three standard sets of handcrafted features.
Oussama Zayene, Nadia Hajjej, Sameh Masmoudi Touj, Soumaya Ben Mansour, Jean Hennebert, Rolf Ingold, Najoua Essoukri Ben Amara
This paper describes the AcTiVComp: detection and recognition of Arabic Text in Video competition in conjunction with the 23rd International Conference on Pattern Recognition (ICPR). The main objective of this competition is to evaluate the performance of participants' algorithms to automatically locate and/or recognize overlay text lines in Arabic video frames using the freely available AcTiV dataset. In this first edition of AcTiVComp, four groups with five systems are participating to the competition. In the detection challenge, the systems are compared based on the standard assessment metrics (i.e. recall, precision and F-score). The recognition results evaluation is based on the recognition rates at the character, word and line levels. The systems were tested in a blind manner on the closed-test set of the AcTiV dataset which is unknown to all participants. In addition to the test results, we also provide a short description of the participating groups and their systems.
Kai Chen, Mathias Seuret, Marcus Liwicki, Jean Hennebert, Cheng-Lin Liu, Rolf Ingold
Proceedings of the 2016 15th International Conference on Frontiers in Handwriting Recognition (ICFHR), 23-26 October 2016, Shenzhen, China
In this paper, we present a Conditional Random Field (CRF) model to deal with the problem of segmenting handwritten historical document images into different regions. We consider page segmentation as a pixel-labeling problem, i.e., each pixel is assigned to one of a set of labels. Features are learned from pixel intensity values with stacked convolutional autoencoders in an unsupervised manner. The features are used for the purpose of initial classification with a multilayer perceptron. Then a CRF model is introduced for modeling the local and contextual information jointly in order to improve the segmentation. For the purpose of decreasing the time complexity, we perform labeling at superpixel level. In the CRF model, graph nodes are represented by superpixels. The label of each pixel is determined by the label of the superpixel to which it belongs. Experiments on three public datasets demonstrate that, compared to previous methods, the proposed method achieves more accurate segmentation results and is much faster.
Pascal Wicht, Andreas Fischer, Jean Hennebert
Proceedings of the 25th International Conference on Artificial Neural Networks and Machine Learning (ICANN), 6-9 September 2016, Barcelona, Spain
To spot keywords on handwritten documents, we present a hybrid keyword spotting system, based on features extracted with Convolutional Deep Belief Networks and using Dynamic Time Warping for word scoring. Features are learned from word images, in an unsupervised manner, using a sliding window to extract horizontal patches. For two single writer historical data sets, it is shown that the proposed learned feature extractor outperforms two standard sets of features.
Proceedings of the 7th IAPR TC3 Workshop, Artificial Neural Networks in Pattern Recognition (ANNPR) 2016, 28-30 September 2016, Ulm Germany
Although Graphics Processing Units (GPUs) seem to currently be the best platform to train machine learning models, most research laboratories are still only equipped with standard CPU systems. In this paper, we investigate multiple techniques to speedup the training of Restricted Boltzmann Machine (RBM) models and Convolutional RBM (CRBM) models on CPU with the Contrastive Divergence (CD) algorithm. Experimentally, we show that the proposed techniques can reduce the training time by up to 30 times for RBM and up to 12 times for CRBM, on a data set of handwritten digits.
Kai Chen, Cheng-Lin Liu, Mathias Seuret, Marcus Liwicki, Jean Hennebert, Rolf Ingold
Proceedings of the 2016 12th IAPR Workshop on Document Analysis Systems (DAS), 11-14 April 2016, Santorini, Greece
In this paper, we present an efficient page segmentation method for historical document images. Many existing methods either rely on hand-crafted features or perform rather slow as they treat the problem as a pixel-level assignment problem. In order to create a feasible method for real applications, we propose to use superpixels as basic units of segmentation, and features are learned directly from pixels. An image is first oversegmented into superpixels with the simple linear iterative clustering (SLIC) algorithm. Then, each superpixel is represented by the features of its central pixel. The features are learned from pixel intensity values with stacked convolutional autoencoders in an unsupervised manner. A support vector machine (SVM) classifier is used to classify superpixels into four classes: periphery, background, text block, and decoration. Finally, the segmentation results are refined by a connected component based smoothing procedure. Experiments on three public datasets demonstrate that compared to our previous method, the proposed method is much faster and achieves comparable segmentation results. Additionally, much fewer pixels are used for classifier training.
Oussama Zayene, Mathias Seuret, Sameh M. Touj, Jean Hennebert, Rolf Ingold, Najoua Essoukri Ben Amara
Text detection in videos is a challenging problem due to variety of text specificities, presence of complex background and anti-aliasing/compression artifacts. In this paper, we present an approach for horizontally aligned artificial text detection in Arabic news video. The novelty of this method revolves around the combination of two techniques: an adapted version of the Stroke Width Transform (SWT) algorithm and a convolutional auto-encoder (CAE). First, the SWT extracts text candidates' components. They are then filtered and grouped using geometric constraints and Stroke Width information. Second, the CAE is used as an unsupervised feature learning method to discriminate the obtained textline candidates as text or non-text. We assess the proposed approach on the public Arabic-Text-in-Video database (AcTiV-DB) using different evaluation protocols including data from several TV channels. Experiments indicate that the use of learned features significantly improves the text detection results.
Proceedings of 2016 2nd International Conference on Advanced Technologies for Signal and Image Processing (ATSIP), 21-23 March 2016, Monastir, Tunisia
Benchmark datasets and their corresponding evaluation protocols are commonly used by the computer vision community, in a variety of application domains, to assess the performance of existing systems. Even though text detection and recognition in video has seen much progress in recent years, relatively little work has been done to propose standardized annotations and evaluation protocols especially for Arabic Video-OCR systems. In this paper, we present a framework for evaluating text detection in videos. Additionally, dataset, ground-truth annotations and evaluation protocols, are provided for Arabic text detection. Moreover, two published text detection algorithms are tested on a part of the AcTiV database and evaluated using a set of the proposed evaluation protocols.
Olivier Liechti, Laurent Prévost, Valentin Delaye, Jean Hennebert, Vincent Grivel, Jean-Philippe Rey, Jonathan Depraz, Marc Sommer
Proceedings of WoT '15: Proceedings of the 6th International Workshop on the Web of Things, 26-28 October 2015, Seoul, South Korea
This paper presents the iFLUX middleware, designed to provide a lightweight integration solution for Smart City applications. Based on three core abstractions, namely event sources, action targets and rules, iFLUX makes it very easy to expose sensors and actuators through REST APIs so that they can be integrated in application-level workflows. Sensors and actuators can be smart objects integrating hardware and software, but can also be pure software services. In the paper, we introduce the iFLUX programming model and describe how it has been implemented in a middleware platform. We also report on how the platform has been used and evaluated in various contexts. While iFLUX has been initially designed in the context of Smart City applications, it is generic and applicable to other domains where hardware and software components are connected through the Web.
Antonio Ridi, Nikos Zarkadis, Christophe Gisler, Jean Hennebert
Proceedings of 2015 IEEE International Conference on Data Science and Advanced Analytics (DSAA), 19-21 October 2015, Paris, France
Activity recognition and prediction in buildings can have multiple positive effects in buildings: improve elderly monitoring, detect intrusions, maximize energy savings and optimize occupant comfort. In this paper we apply human activity recognition by using data coming from a network of motion and door sensors distributed in a Smart Home environment. We use Hidden Markov Models (HMM) as the basis of a machine learning algorithm on data collected over an 8-month period from a single-occupant home available as part of the WSU CASAS Smart Home project. In the first implementation the HMM models 24 hours of activities and classifies them in 8 distinct activity categories with an accuracy rate of 84.6%. To improve the identification rate and to help detect potential abnormalities related with the duration of an activity (i.e. when certain activities last too much), we implement minimum duration modeling where the algorithm is forced to remain in a certain state for a specific amount of time. Two subsequent implementations of the minimum duration HMM (mean-based length modeling and quantile length modeling) yield a further 2% improvement of the identification rate. To predict the sequence of activities in the future, Artificial Neural Networks (ANN) are employed and identified activities clustered in 3 principal activity groups with an average accuracy rate of 71-77.5%, depending on the forecasting window. To explore the energy savings potential, we apply thermal dynamic simulations on buildings in central European climate for a period of 65 days during the winter and we obtain energy savings for space heating of up to 17% with 3-hour forecasting for two different types of buildings.
Baptiste Wicht, Jean Hennebert
Proceedings of 2015 13th International Conference on Document Analysis and Recognition (ICDAR), 23-26 August 2015, Tunis, Tunisia
In this paper, we propose a method to recognize Sudoku puzzles containing both handwritten and printed digits from images taken with a mobile camera. The grid and the digits are detected using various image processing techniques including Hough Transform and Contour Detection. A Convolutional Deep Belief Network is then used to extract high-level features from raw pixels. The features are finally classified using a Support Vector Machine. One of the scientific question addressed here is about the capability of the Deep Belief Network to learn extracting features on mixed inputs, printed and handwritten. The system is thoroughly tested on a set of 200 Sudoku images captured with smartphone cameras under varying conditions, e.g. distortion and shadows. The system shows promising results with 92% of the cells correctly classified. When cell detection errors are not taken into account, the cell recognition accuracy increases to 97.7%. Interestingly, the Deep Belief Network is able to handle the complex conditions often present on images taken with phone cameras and the complexity of mixed printed and handwritten digits.
Oussama Zayene, Jean Hennebert, Sameh Masmoudi Touj, Rolf Ingold, Najoua Essoukri Ben Amara
Proceedings of the 2015 13th International Conference on Document Analysis and Recognition (ICDAR), 23-26 August 2015, Tunis, Tunisia
Recently, promising results have been reported on video text detection and recognition. Most of the proposed methods are tested on private datasets with non-uniform evaluation metrics. We report here on the development of a publicly accessible annotated video dataset designed to assess the performance of different artificial Arabic text detection, tracking and recognition systems. The dataset includes 80 videos (more than 850,000 frames) collected from 4 different Arabic news channels. An attempt was made to ensure maximum diversities of the textual content in terms of size, position and background. This data is accompanied by detailed annotations for each textbox. We also present a region-based text detection approach in addition to a set of evaluation protocols on which the performance of different systems can be measured.
Kai Chen, Mathias Seuret, Marcus Liwicki, Jean Hennebert, Rolf Ingold
In this paper, we present an unsupervised feature learning method for page segmentation of historical handwritten documents available as color images. We consider page segmentation as a pixel labeling problem, i.e., each pixel is classified as either periphery, background, text block, or decoration. Traditional methods in this area rely on carefully hand-crafted features or large amounts of prior knowledge. In contrast, we apply convolutional autoencoders to learn features directly from pixel intensity values. Then, using these features to train an SVM, we achieve high quality segmentation without any assumption of specific topologies and shapes. Experiments on three public datasets demonstrate the effectiveness and superiority of the proposed approach.
Proceedings of 2015 IEEE International Conference on Pervasive Computing and Communication Workshops (PerCom Workshops), 23-27 March 2015, St. Louis, MO, USA
In this paper we assess about the recognition of User Interaction events when handling electrical devices. This work is placed in the context of Intrusive Load Monitoring used for appliance recognition. ILM implies several Smart Metering Sensors to be placed inside the environment under analysis (in our case we have one Smart Metering Sensor per device). Our existing system is able to recognise the appliance class (as coffee machine, printer, etc.) and the sequence of states (typically Active / Non-Active) by using Hidden Markov Models as machine learning algorithm. In this paper we add a new layer to our system architecture called User Interaction Layer, aimed to infer the moments (called User Interaction events) during which the user interacts with the appliance. This layer uses as input the information coming from HMM (i.e. the recognised appliance class and the sequence of states). The User Interaction events are derived from the analysis of the transitions in the sequences of states and a ruled-based system adds or removes these events depending on the recognised class. Finally we compare the list of events with the ground truth and we obtain three different accuracy rates: (i) 96.3% when the correct model and the real sequence of states are known a priori, (ii) 82.5% when only the correct model is known and (iii) 80.5% with no a priori information.
Proceedings of 2015 IEEE Wireless Communications and Networking Conference Workshops (WCNCW), 9-12 March 2015, New Orleans, LA, USA
The automatic identification of appliances through the analysis of their electricity consumption has several purposes in Smart Buildings including better understanding of the energy consumption, appliance maintenance and indirect observation of human activities. Electric signatures are typically acquired with IoT smart plugs integrated or added to wall sockets. We observe an increasing number of research teams working on this topic under the umbrella Intrusive Load Monitoring. This term is used as opposition to Non-Intrusive Load Monitoring that refers to the use of global smart meters. We first present the latest evolutions of the ACS-F database, a collections of signatures that we made available for the scientific community. The database contains different brands and/or models of appliances with up to 450 signatures. Two evaluation protocols are provided with the database to benchmark systems able to recognise appliances from their electric signature. We present in this paper two additional evaluation protocols intended to measure the impact of the analysis window length. Finally, we present our current best results using machine learning approaches on the 4 evaluation protocols.
Christophe Gisler, Antonio Ridi, Milène Fauquex, Dominique Genoud, Jean Hennebert
Proceedings of the 6th International Conference of Soft Computing and Pattern Recognition (SoCPaR), 11-14 August 2014, Tunis, Tunisia
Diagnosing the glaucoma is a very difficult task for healthcare professionals. High intraocular pressure (IOP) remains the main treatable symptom of this degenerative disease which leads to blindness. Nowadays, new types of wearable sensors, such as the contact lens sensor Triggerfish ® , provide an automated recording of 24-hour profile of ocular dimensional changes related to IOP. Through several clinical studies, more and more IOP-related profiles have been recorded by those sensors and made available for elaborating data-driven experiments. The objective of such experiments is to analyse and detect IOP pattern differences between ill and healthy subjects. The potential is to provide medical doctors with analysis and detection tools allowing them to better diagnose and treat glaucoma. In this paper we present the methodologies, signal processing and machine learning algorithms elaborated in the task of automated detection of glaucomatous IOP-related profiles within a set of 100 24-hour recordings. As first convincing results, we obtained a classification ROC AUC of 81.5%.
Gérôme Bovet, Gautier Briard, Jean Hennebert
Proceedings of the WoT '14: Proceedings of the 5th International Workshop on Web of Things, 8 October 2014, Cambridge, MA, USA
Data storage has become a major topic in sensor networks as large quantities of data need to be archived for future processing. In this paper, we present a cloud storage solution benefiting from the available memory on smart things becoming data nodes. In-network storage reduces the heavy traffic resulting of the transmission of all the data to an outside central sink. The system built on agents allows an autonomous management of the cloud and therefore requires no human in the loop. It also makes an intensive use of Web technologies to follow the clear trend of sensors adopting the Web-of-Things paradigm. Further, we make a performance evaluation demonstrating its suitability in building management systems.
Proceedings of the 2014 International Conference on Data Science and Advanced Analytics (DSAA), 30 October - 1 November 2014, Shanghai, China
We asset about the analysis of electrical appliance consumption signatures for the identification task. We apply Hidden Markov Models to appliance signatures for the identification of their category and of the most probable sequence of states. The electrical signatures are measured at low frequency (10 -1 Hz) and are sourced from a specific database. We follow two predefined protocols for providing comparable results. Recovering information on the actual appliance state permits to potentially adopt energy saving measures, as switching off stand-by appliances or, generally speaking, changing their state. Moreover, in most of the cases appliance states are related to user activities: the user interaction usually involves a transition of the appliance state. Information about the state transition could be useful in Smart Home / Building Systems to reduce energy consumption and increase human comfort.We report the results of the classification tasks in terms of confusion matrices and accuracy rates. Finally, we present our application for a real-time data visualization and the recognition of the appliance category with its actual state.
Internet-of-Things (IoT) devices, especially sensors are producing large quantities of data that can be used for gathering knowledge. In this field, machine learning technologies are increasingly used to build versatile data-driven models. In this paper, we present a novel architecture able to execute machine learning algorithms within the sensor network, presenting advantages in terms of privacy and data transfer efficiency. We first argument that some classes of machine learning algorithms are compatible with this approach, namely based on the use of generative models that allow a distribution of the computation on a set of nodes. We then detail our architecture proposal, leveraging on the use of Web-of-Things technologies to ease integration into networks. The convergence of machine learning generative models and Web-of-Things paradigms leads us to the concept of virtual things exposing higher level knowledge by exploiting sensor data in the network. Finally, we demonstrate with a real scenario the feasibility and performances of our proposal.
Proceedings of the 4th International Conference on Image Processing Theory, Tools and Applications (IPTA) 2014, 14-17 October 2014, Paris, France
In this paper, we present a semi-automatic news video annotation tool. The tool and its algorithms are dedicated to artificial Arabic text embedded in video news in the form of static text as well as scrolling one. It is performed at two different levels. Including specificities of Arabic script, the tool manages a global level which concerns the entire video and a local level which concerns any specific frame extracted from the video. The global annotation is performed manually thanks to a user interface. As a result of this step, we obtain the global xml file. The local annotation at the frame level is done automatically according to the information contained in the global metafile and a proposed text tracking algorithm. The main application of our tool is the ground truthing of textual information in video content. It is being used for this purpose in the Arabic Text in Video (AcTiV) database project in our lab. One of the functions that AcTiV provides, is a benchmark to compare existing and future Arabic video OCR systems.
Proceedings of the 4th International Conference on the Internet of Things, 6-8 October 2014, Cambridge, MA, USA
Machine Learning (ML) approaches are increasingly used to model data coming from sensor networks. Typical ML implementations are cpu intensive and are often running server-side. However, IoT devices provide increasing cpu capabilities and some classes of ML algorithms are compatible with distribution and downward scalability. In this demonstration we explore the possibility of distributing ML tasks to IoT devices in the sensor network. We demonstrate a concrete scenario of appliance recognition where a smart plug provides electrical measures that are distributed to WiFi nodes running the ML algorithms. Each node estimates class-conditional probabilities that are then merged for recognizing the appliance category. Finally, our architectures relies on Web technologies for complying with Web-of-Things paradigms.
Kai Chen, Hao Wei, Jean Hennebert, Rolf Ingold, Marcus Liwicki
Proceedings of the 2014 14th International Conference on Frontiers in Handwriting Recognition (ICFHR), 1-4 September 2014, Hersonissos, Greece
In this paper we present a physical structure detection method for historical handwritten document images. We considered layout analysis as a pixel labeling problem. By classifying each pixel as either periphery, background, text block, or decoration, we achieve high quality segmentation without any assumption of specific topologies and shapes. Various color and texture features such as color variance, smoothness, Laplacian, Local Binary Patterns, and Gabor Dominant Orientation Histogram are used for classification. Some of these features have so far not got many attentions for document image layout analysis. By applying an Improved Fast Correlation-Based Filter feature selection algorithm, the redundant and irrelevant features are removed. Finally, the segmentation results are refined by a smoothing post-processing procedure. The proposed method is demonstrated by experiments conducted on three different historical handwritten document image datasets. Experiments show the benefit of combining various color and texture features for classification. The results also show the advantage of using a feature selection method to choose optimal feature subset. By applying the proposed method we achieve superior accuracy compared with earlier work on several datasets, e.g., We achieved 93% accuracy compared with 91% of the previous method on the Parzival dataset which contains about 100 million pixels.
Kai Chen, Hao Wei, Marcus Liwicki, Jean Hennebert, Rolf Ingold
Proceedings of the 22nd International Conference on Pattern Recognition, 24-28 August 2014, Stockholm, Sweden
In this paper we present a novel text line segmentation method for historical manuscript images. We use a pyramidal approach where at the first level, pixels are classified into: text, background, decoration, and out of page, at the second level, text regions are split into text line and non text line. Color and texture features based on Local Binary Patterns and Gabor Dominant Orientation are used for classification. By applying a modified Fast Correlation-Based Filter feature selection algorithm, redundant and irrelevant features are removed. Finally, the text line segmentation results are refined by a smoothing post-processing procedure. Unlike other projection profile or connected components methods, the proposed algorithm does not use any script-specific knowledge and is applicable to color images. The proposed algorithm is evaluated on three historical manuscript image datasets of diverse nature and achieved an average precision of 91% and recall of 84%. Experiments also show that the proposed algorithm is robust with respect to changes of the writing style, page layout, and noise on the image.
In this paper, we propose a method to detect and recognize a Sudoku puzzle on images taken from a mobile camera. The lines of the grid are detected with a Hough transform. The grid is then recomposed from the lines. The digits position are extracted from the grid and finally, each character is recognized using a Deep Belief Network (DBN). To test our implementation, we collected and made public a dataset of Sudoku images coming from cell phones. Our method proved successful on our dataset, achieving 87.5% of correct detection on the testing set. Only 0.37% of the cells were incorrectly guessed. The algorithm is capable of handling some alterations of the images, often present on phone-based images, such as distortion, perspective, shadows, illumination gradients or scaling. On average, our solution is able to produce a result from a Sudoku in less than 100ms.
Proceedings of the 2014 6th International Conference of Soft Computing and Pattern Recognition (SoCPaR), 11-14 August 2014, Tunis, Tunisia
We present ACS-F2, a new electric consumption signature database acquired from domestic appliances. The scenario of use is appliance identification with emerging applications such as domestic electricity consumption understanding, load shedding management and indirect human activity monitoring. The novelty of our work is to use low-end electricity consumption sensors typically located at the plug. Our approach consists in acquiring signatures at a low frequency, which contrast with high frequency transient analysis approaches that are costlier and have been well studied in former research works. Electrical consumption signatures comprise real power, reactive power, RMS current, RMS voltage, frequency and phase of voltage relative to current. A total of 225 appliances were recorded over two sessions of one hour. The database is balanced with 15 different brands/models spread into 15 categories. Two realistic appliance recognition protocols are proposed and the database is made freely available to the scientific community for the experiment reproducibility. We also report on recognition results following these protocols and using baseline recognition algorithms like k-NN and GMM.
Electricity load monitoring of appliances has become an important task considering the recent economic and ecological trends. In this game, machine learning has an important part to play, allowing for energy consumption understanding, critical equipment monitoring and even human activity recognition. This paper provides a survey of current researches on Intrusive Load Monitoring (ILM) techniques. ILM relies on low-end electricity meter devices spread inside the habitations, as opposed to Non-Intrusive Load Monitoring (NILM) that relies on an unique point of measurement, the smart meter. Potential applications and principles of ILMs are presented and compared to NILM. A focus is also given on feature extraction and machine learning algorithms typically used for ILM applications.
Antonio Ridi, Jean Hennebert
Proceedings of the International Workshop on Enabling ICT for Smart Buildings (ICT-SB 2014), 2-5 June 2014, Hasselt, Belgium ; Procedia Computer Science
The automatic recognition of appliances through the monitoring of their electricity consumption finds many applications in smart buildings. In this paper we discuss the use of Hidden Markov Models (HMMs) for appliance recognition using so-called intrusive load monitoring (ILM) devices. Our motivation is found in the observation of electric signatures of appliances that usually show time varying profiles depending to the use made of the appliance or to the intrinsic internal operating of the appliance. To determine the benefit of such modelling, we propose a comparison of stateless modelling based on Gaussian mixture models and state-based models using Hidden Markov Models. The comparison is run on the publicly available database ACS-F1. We also compare differ- ent approaches to determine the best model topologies. More specifically we compare the use of a priori information on the device, a procedure based on a criteria of log-likelihood maximization and a heuristic approach.
Kai Chen, Jean Hennebert
Proceedings of the 6th Mexican Conference, MCPR 2014, Pattern Recognition, 25-28 June 2014, Cancun, Mexico
We present the evaluation of a product identification task using the LIRe system and SURF (Speeded-Up Robust Features) for content-based image retrieval (CBIR). The evaluation is performed on the Fribourg Product Image Database (FPID) that contains more than 3’000 pictures of consumer products taken using mobile phone cameras in realistic conditions. Using the evaluation protocol proposed with FPID, we explore the performance of different preprocessing and feature extraction. We observe that by using SURF, we can improve significantly the performance on this task. Image resizing and Lucene indexing are used in order to speed up CBIR task with SURF. We also show the benefit of using simple preprocessing of the images such as a proportional cropping of the images. The experiments demonstrate the effectiveness of the proposed method for the product identification task.
Gérôme Bovet, Jean Hennebert
Procedia Computer Science ; Proceedings of the International Workshop on Enabling ICT for Smart Buildings (ICT-SB 2014), 2-5 June 2014, Hasselt, Belgium
Offices, factories and even private housings are more and more endowed with building management systems (BMS) targeting an increase of comfort as well as lowering energy costs. This expansion is made possible by the progress realized in pervasive computing, providing small sized and affordable sensing devices. However, current BMS are often based on proprietary tech- nologies, making their interoperability and evolution more difficult. For example, we observe the emergence of new applications based on intelligent data analysis able to compute more complex models about the use of the building. Such applications rely on heterogeneous sets of sensors, web data, user feedback and self-learning algorithms. In this position paper, we discuss the role of Web technologies for standardizing the application layer, and thus providing a framework for developing advanced building applications. We present our vision of TASSo, a layered Web model facing actual and future challenges for building management systems.
Proceedings of the 3rd IEEE international workshop on the IoT: Smart Objects and Services, 19 June 2014, Sydney, NSW, Australia
Nowadays, pervasive application scenarios relying on sensor networks are gaining momentum. The field of smart buildings is a promising playground where the use of sensors allows a reduction of the overall energy consumption. Most of current applications are using the classical DNS which is not suited for the Internet-of-Things because of requiring humans to get it working. From another perspective, Web technologies are pushing in sensor networks following the Web-of-Things paradigm advocating to use RESTful APIs for manipulating resources representing device capabilities. Being aware of these two observations, we propose to build on top of Web technologies leading to a novel naming system that is entirely autonomous. In this work, we describe the architecture supporting what can be called an autonomous Web-oriented naming system. As proof of concept, we simulate a rather large building and compare the behaviour of our approach to the legacy DNS and Multicast DNS (mDNS).
Gerome Bovet, Jean Hennebert
Proceedings of the 2014 6th International Conference on New Technologies, Mobility and Security (NTMS), 30 March-2 April 2014, Dubai, United Arab Emirates
Nowadays, our surrounding environment is more and more scattered with various types of sensors. Due to their intrinsic properties and representation formats, they form small islands isolated from each other. In order to increase interoperability and release their full capabilities, we propose to represent devices descriptions including data and service invocation with a common model allowing to compose mashups of heterogeneous sensors. Pushing this paradigm further, we also propose to augment service descriptions with a discovery protocol easing automatic assimilation of knowledge. In this work, we describe the architecture supporting what can be called a Semantic Sensor Web-of-Things. As proof of concept, we apply our proposal to the domain of smart buildings, composing a novel ontology covering heterogeneous sensing, actuation and service invocation. Our architecture also emphasizes on the energetic aspect and is optimized for constrained environments.
Nayla Sokhn, Richard Baltensperger, Louis-Felix Bersier, Ulrich-Ultes Nitsche, Jean Hennebert
Proceedings of the 2013 International Conference on Signal-Image Technology & Internet-Based Systems, 2-5 December 2013, Kyoto, Japan
The structure of networks has always been interesting for researchers. Investigating their unique architecture allows to capture insights and to understand the function and evolution of these complex systems. Ecological networks such as food-webs and niche-overlap graphs are considered as complex systems. The main purpose of this work is to compare the topology of 15 real niche-overlap graphs with random ones. Five measures are treated in this study: (1) the clustering coefficient, (2) the between ness centrality, (3) the assortativity coefficient, (4) the modularity and (5) the number of chord less cycles. Significant differences between real and random networks are observed. Firstly, we show that niche-overlap graphs display a higher clustering and a higher modularity compared to random networks. Moreover we find that random networks have barely nodes that belong to a unique sub graph (i.e. between ness centrality equal to 0) and highlight the presence of a small number of chord less cycles compared to real networks. These analyses may provide new insights in the structure of these real niche-overlap graphs and may give important implications on the functional organization of species competing for some resources and on the dynamics of these systems.
Proceedings of the 7th FTRA International Conference on Multimedia and Ubiquitous Engineering (MUE 2013), 9-11 May 2013, Daegu, South Korea
Leveraging on the Web-of-Things (WoT) allows standardizing the access of things from an application level point of view. The protocols of the Web and especially HTTP are offering new ways to build mashups of things consisting of sensors and actuators. Two communication protocols are now emerging in the WoT domain for event-based data exchang, namely WebSockets and RESTful APIs. In this work, we motivate and demonstrate the use of a hybrid layer able to choose dynamically the most energy efficient protocol.