Search

Segmentation is a fundamental problem but not the ultimate goal, it is a stepping stone to higher level application problems. It consists in associating each of the low-level image pixels to the class they locally represent. This task completes image analysis tasks such as visual scene classification and instance level object detection. It enables high level applications in a variety of domains, from images and video indexing to autonomous vehicle driving and medical image analysis. Recently, deep learning approaches have pushed image segmentation and object instance segmentation in a new era with impressive performance levels. However, several challenges have to be faced to train those approaches in an effective way for each of the case studies, dealing with few training samples, specific data, strong target imbalance and so on. This chapter reviews the image segmentation task and recent advanced strategies to face those potential issues in a variety of application domains. Current challenges to address are highlighted for future research directions.

International audience; The present paper introduces convolution and pooling operators for indexed images. These operators can be used on images that do not provide Cartesian grids of pixels, as long as a list of neighbor's indices can be provided for each pixel. They are foreseen being useful for convolutional neural networks (CNN) applied to special sensors, especially in science, without requiring image pre-processing. The present work explains the method and its implementation in the Pytorch framework and shows an application of the indexed kernels to the classification task of images with hexagonal lattices using CNN. The obtained results show that the method gives the same performances as the standard convolution kernels. Indexed convolution thus makes deep neural network frameworks more general and capable of addressing unconventional image lattices. The current implementation, as well as code to reproduce the experiments described in this paper are made available as open-source resources on the repository www.github.com/IndexedConv.

International audience; We compare a recent dehazing method based on deep learning , Dehazenet, with traditional state-of-the-art approach, on benchmark data with reference. Dehazenet estimates the depth map from a single color image, which is used to inverse the Koschmieder model of imaging in the presence of haze. In this sense, the solution is still attached to the Koschmieder model. We demonstrate that this method exhibits the same limitation than other inversions of this imaging model.

International audience; In our data driven world, categorization is of major importance to help end-users and decision makers understanding information structures. Supervised learning techniques rely on annotated samples that are often difficult to obtain and training often overfits. On the other hand, unsupervised clustering techniques study the structure of the data without disposing of any training data. Given the difficulty of the task, supervised learning often outperforms unsupervised learning. A compromise is to use a partial knowledge, selected in a smart way, in order to boost performance while minimizing learning costs, what is called semi-supervised learning. In such use case, Spectral Clustering proved to be an efficient method. Also, Deep Learning outperformed several state of the art classification approaches and it is interesting to test it in our context. In this paper, we firstly introduce the concept of Deep Learning into an active semi-supervised clustering process and compare it with Spectral Clustering. Secondly, we introduce constraint propagation and demonstrate how it maximizes partitioning quality while reducing annotation costs. Experimental validation is conducted on two different real datasets. Results show the potential of the clustering methods.

This work addresses the automatic reading of sale receipts acquired with a smartphone's camera and the extraction of essential informations like the store brands, all the purchased products , their price etc. It is divided into two major tasks : the optical character recognition that is made complex due to the nature of the document often damaged, crumpled or torned and the semantic data analysis to identify purchased product without ambiguity. In this paper, we introduce a solution that enables a "guided" capture on smartphone and a receipt decoding on remote server., L'objectif de ces travaux est de créer un système capable, à partir d'une simple vue prise par un smartphone, d'exploiter le contenu d'un ticket de caisse, et d'en extraire des informations telles que le point de vente, les produits ache-tés, leurs prix etc. On distingue deux étapes majeures : la reconnaissance optique de caractères sur des tickets qui peuvent être froissés, déchirés... et l'analyse sémantique afin d'identifier les produits achetés sans ambiguïté. Dans ce papier, nous présentons une solution qui effectue une acquisition "guidée" d'une image sur smartphone et une lecture du contenu sur un serveur distant.

International audience; The paper explores how multimedia approaches used in image understanding tasks could be adapted and used in remote sensing image analysis. Two approaches are investigated: the classical Bag of Visual Words (BoVW) approach and the Deep Learning approach. Tests are performed for the classification of the UC Merced Land Use Dataset which provide better results than the state of the art.

International audience; In this paper, we make the link between biological knowledge of the Human Visual System (HVS) and computer vision processing. The goal is to show the advantage of using biological perception models in all day life computer vision applications. On the one hand, biological researches on the HVS allow the exploration of the physiological properties of Human to see, to understand, to learn and to adapt to the environment. The human retina and the primary visual cortex V1 have been one of the main explored parts of the HVS. This enthusiastic topic gives a better understanding of the different cells networks involved from a biological point of view and models their effect from a signal processing point of view. Some models have been proposed [1, 2, 3] and have already shown interesting human-like vision perception properties. The main properties are linked to the adaptivity of the HVS which allows to " see " with a good reliability in various conditions from sunlight to shadow. Real-time retina models allow imagining their use in today's vision applications needs and some of them already show the potential such as artificial retina [4], image enhancement [5]. On the other hand, computer vision is a wide research area in which several mathematical approaches are used in order to solve visual perception problems. Various research axis is explored such as objects recognition and tracking, face analysis, visual scene exploration, etc... All these computer vision tasks are in general achieved easily by humans. Nevertheless, they are solved with complex mathematical approaches which are not linked to human perception. The proposed algorithms are often sensitive to environment changes (luminance variations, acquisition noise etc.) and require complex setup stages in order to obtain good performances in specific conditions. We propose here to show how it is possible to use the efficient real-time retina model proposed in [4] and the V1 primary cortex model of [6] to achieve real-time computer vision processing. The spatio-temporal properties of the retina model allow a robust spatio-temporal contours extraction by modelling the parvocellular (details) and magnocellular (motion) channels which are robust against noise and luminance variations. These two channels are the basis of the computer vision algorithms that we propose. Coupled with the V1 cortex model which allows object and shape recognition, we propose here application examples: A face analysis system which detects the open or closed eyes and mouth states by using some energetic properties of the studied face features allowing hypo-vigilance detection for drivers; A moving object segmentation and tracking system based on the description of the motion energy generated by the moving objects; And finally an object classifier based on the object texture and structure analysis. All these applications are exclusively based on a bio-inspired architecture. They show the interest to involve the knowledge of the HSV in order to solve computer vision applications. Several other computer vision applications can exploit such a philosophy opening a new way to solve vision problems.

International audience; In the computer vision field, semantic segmentation represents a very interesting task. Convolutional Neural Network methods have shown their great performances in comparison with other semantic segmentation methods. In this paper, we propose a multiscale fully convolutional DenseNet approach for semantic segmentation. Our approach is based on the successful fully convolutional DenseNet method. It is reinforced by integrating a multiscale kernel prediction after the last dense block which performs model averaging over different spatial scales and provides more flexibility of our network to presume more information. Experiments on two semantic segmentation benchmarks: CamVid and Cityscapes have shown the effectiveness of our approach which has outperformed many recent works.

National audience