Reda Belaiche

Les technologies de l'interaction homme-machine se concentrent de plus en plus sur l'être humain, que ce soit sur son identité, ou bien sur son état physique et mental. Des progrès conséquents ont été réalisés depuis quelques décennies. Cependant l'étude des pensées et des émotions reste encore un domaine peu développé, mais qui a commencé à grandement gagner en intérêt. Dans ce domaine, l'analyse des expressions faciales est le traitement préférentiel.Contrairement à une macro-expression, visible à l'œil, une micro-expression est un type d'expression faciale involontaire extrêmement rapide et de très faible intensité. La communauté scientifique en vision par ordinateur étudie depuis quelques années les façons de reconnaître automatiquement les micro-expressions à l'aide de cameras rapides et de programmes informatiques. Il s'agit néanmoins d'un problème difficile, de par la nature de ces micro-expressions.À partir des dernières avancées techniques en machine learning, nous avons proposé plusieurs traitements de reconnaissance des micro-expressions sur des vidéos tout en mettant l'accent sur la rapidité d'exécution et les faibles besoins en mémoire. Suivant l'évolution de la littérature, nous avons dans un premier temps proposé des améliorations quant à l'utilisation de descripteurs comme le LBP-TOP, puis nous avons configuré la structure d'un réseau de neurones convolutifs pour le rendre plus compact sans perdre en précision.Avec des résultats similaires aux méthodes actuelles, nous avons considérablement réduit les besoins en mémoire et en calculs. Nous avons finalement démontré la faisabilité de l'exécution en conditions réelles d'un système complet réalisant à la fois la détection et la reconnaissance des micro-expressions.

Micro-Expression (ME) recognition is a hot topic in computer vision as it presents a gateway to capture and understand daily human emotions. It is nonetheless a challenging problem due to ME typically being transient (lasting less than 200 ms) and subtle. Recent advances in machine learning enable new and effective methods to be adopted for solving diverse computer vision tasks. In particular, the use of deep learning techniques on large datasets outperforms classical approaches based on classical machine learning which rely on hand-crafted features. Even though available datasets for spontaneous ME are scarce and much smaller, using off-the-shelf Convolutional Neural Networks (CNNs) still demonstrates satisfactory classification results. However, these networks are intense in terms of memory consumption and computational resources. This poses great challenges when deploying CNN-based solutions in many applications, such as driver monitoring and comprehension recognition in virtual classrooms, which demand fast and accurate recognition. As these networks were initially designed for tasks of different domains, they are over-parameterized and need to be optimized for ME recognition. In this paper, we propose a new network based on the well-known ResNet18 which we optimized for ME classification in two ways. Firstly, we reduced the depth of the network by removing residual layers. Secondly, we introduced a more compact representation of optical flow used as input to the network. We present extensive experiments and demonstrate that the proposed network obtains accuracies comparable to the state-of-the-art methods while significantly reducing the necessary memory space. Our best classification accuracy was 60.17% on the challenging composite dataset containing five objectives classes. Our method takes only 24.6 ms for classifying a ME video clip (less than the occurrence time of the shortest ME which lasts 40 ms). Our CNN design is suitable for real-time embedded applications with limited memory and computing resources.

Machine learning has known a tremendous growth within the last years, and lately, thanks to that, some computer vision algorithms started to access what is difficult or even impossible to perceive by the human eye. While deep learning based computer vision algorithms have made themselves more and more present in the recent years, more classical feature extraction methods, such as the ones based on Local Binary Patterns (LBP), still present a non negligible interest, especially when dealing with small datasets. Furthermore, this operator has proven to be quite useful for facial emotions and human gestures recognition in general. Micro-Expression (ME) classification is among the applications of computer vision that heavily relied on hand crafted features in the past years. LBP Three Orthogonal Planes (LBP_TOP) is one of the most used hand crafted features extractor in the scientific literature to tackle the problem of ME classification. In this paper we present a time unification method that provides better results than the classical LBP_TOP while also drastically reducing the calculations required for feature extraction.