Invited Speakers (ICDIP 2020)

Prof. Yannick Berthoumieu

University of Bordeaux, France

Prof. Yannick Berthoumieu received his Ph.D. degree in signal and image processing in 1996 from the University of Bordeaux (UB). He joined the Bordeaux Engineering School in 1997 as an assistant professor of the Electrical Department. Since 2007 he is a full Professor of the Bordeaux Institute of Technology, where he currently heads the Signal and Image Processing Group (GSI) belonging to the CNRS IMS Laboratory of the University of Bordeaux. His research activities cover areas such as statistical image and video processing, information geometry, as well as probabilistic manifolds in machine learning, classification, segmentation, superresolution and inversion algorithms focused on textural content for optical, radar and hyperspectral imaging applications. He is from 2019 a deputy National Director of the French GdR 720 ISIS (Information, Signal, Image et ViSion) from the « Institut des Sciences Informatiques et de leurs Interactions » of the French National Centre for Scientific Research. Prof. Berthoumieu co-organized international Conferences and Workshops including the IEEE International Conference on Image Processing, the IEEE Image Video and Multidimensional Signal Processing, and the IEEE Statistical Signal Processing. He was the General Chair of the IEEE IVMSP Workshop (2016), and General Chair of the GRETSI conference (2011) both organized in Bordeaux and Publication co-Chair of IEEE SSP Workshop (2005). He is served as area chair for conferences such as IEEE ICIP, ICASSP EUSIPCO or conference on Geometric Science of Information (GSI). He is an Editor Associate of the IEEE Transactions of Image Processing Journal.

Speech Title: Second Order Descriptors and Convolutional Neural Network for Image Classification

Abstract: In the present talk, we will take stock using of second order descriptors in the context of image classification. Indeed, being able to take into account the dependencies between features, covariance matrices have shown a great interest for many practical applications. However, these kind of descriptor sample do not lie in an Euclidean space but a Riemannian manifold. First, in this context, we will realize a overview of mixture-based approaches for encoding a set of covariance matrices. Starting from a Gaussian mixture model (GMM) adapted to the log-Euclidean or affine invariant metric, we propose a Fisher vector descriptor adapted to each of these metrics: the log-Euclidean Fisher Vectors and the Riemannian Fisher Vectors. Some experiments on texture and head pose image classification are conducted to compare these two metrics and to illustrate the potential of these FV-based descriptors compared to state-of-the-art BoW and VLAD-based descriptors. In a second part, we will show how second-order descriptors can be hybridized with convolutional deep learning architecture. Different methods, using ensemble principle and GMM-dictionary coding will be studied. The proposed approaches are integrated in a supervised image classification algorithm based on the encoding of deep neural networks features obtained via a transfer learning approach. For validation and comparison purposes, the proposed approach is tested on various challenging remote sensing datasets compared to very recent methods when considering the first or second layer output of a very deep convolutional neural network.