Aurélie Martin

Contact Information

Email: Address: IRISA Campus Universitaire de Beaulieu 35042 Rennes Cedex - FRANCE Phone: +33 2.99.84.75.51 Fax: +33 2.99.84.71.71 Project Assistant: +33 2.99.84.72.28 (Huguette Béchu) Office: F336 VERT

Profile

2006 - 2009 : PhD Thesis in the TEMICS project (IRISA) on the subject of spatial prediction based on sparse representations, in the context of video compression. PhD in collaboration with THOMSON , in the Content and Delivery Compression group.

2003 - 2006 : Engineering degree from Grenoble Institute of Technology . (School : ENSIEG, energy and signal processing). Third year in telecom department, specialization in telecommunication and transmission systems.

Research Activities

Introduction

Predictive coding was used as the basis of image compression algorithms and it is an important component of MPEG 4 / AVC video coding standard. If the prediction is successful, the energy in the residual error is lower than in the residual image. The residual error can be coded with fewer bits. We focus our work on getting a better intra prediction, thanks to atomic decomposition algorithms.

Sparse representations

The compression efficiency depends on the accuracy of the prediction. However it is not possible to choose a predictor that fits every time to the features in the image. AVC interpolates geometry ; we consider the problem of interpolating textures thanks to predictors that depends on a larger neighborhood.

The goal of sparse representation is to find the best approximation for large classes of signals. The signal Y is represented as a weighted linear combination of elements or atoms from an overcomplete dictionary, A.

The linear representation is the following :

Sparse representation
Linear system : Y = A . X	DCT dictionary	Real DFT dictionary

X is the unknown sparse vector which contains few non-zero coefficients. The issue is to solve this undetermined system of linear equations. There are infinitely many solutions for X. However, under a given constraint (criterion), we can garantee that the sparsest solution is unique. Two types of algorithm can be used to solve this problem :

Criterions
Matching Pursuit (MP)	The Global Matched Filter (GMF)

Intra prediction in H264/AVC

Some frames (I frames) in a video sequence are predicted using blocks contained only within the current frame. This corresponds to the intra frame coding : there is no temporal processing. The frame is divided in blocks, decoded one by one. For the current block to be intra coded, only blocks that have been previously decoded are available. This is the causal neighborhood. The intra prediction is based on the knowledge of this local pixels. The encoder forms a prediction based on propagation or some linear combination of few neighbor pixels. The propagation is performed along few pre-specified directions (horizontal, vertical, diagonal down left, vertical right, ...). This prediction is just able to produce 1D directional interpolation.

Intra prediction based on Sparse Representations

On the contrary of AVC prediction, atomic decomposition makes possible the interpolation of 2D pattern. The dictionary is filled with 2D atoms, the basis functions. For each block, an algorithm find the best set of atoms to modelize the local area formed with neighbor pixels. We assume that the found solution X is sufficient to extrapolate pixels in the current block to be predicted. One challenge is also to find the optimal dictionary, adaptive to the known pixels. If atoms are already very similar to the input signal, the optimal solution would be very sparse.

Results
Original image (512x512)	AVC prediction	AVC + MP prediction

Spatial Inter Layer prediction based on Sparse Representations

In the context of multi layer encoding scheme, the process of prediction exploits dependencies at different resolutions. Each macroblock can be predicted with upsampled lower resolution signal (separable filters). The Scalable Video Coding (SVC) scheme forms a spatial inter layer prediction, added to the AVC intra prediction.

With the technique based on sparse representation, we can exploit two types of correlation at the same time. The prediction is formed thanks to the causal neighborhood in the enhancement layer (higher resolution), but also thanks to upsampled pixels from the base layer (lower resolution). This process can be seen as a prediction refinement : our technique can recover textured features whereas the standard filters only reconstruct smooth areas.

Results
Original image (512x512)	Standard prediction	MP + BL prediction

Publications

	A. Martin , J-J Fuchs , C. Guillemot and D. Thoreau Atomic decomposition dedicated to AVC and spatial SVC prediction ICIP08 15th IEEE International Conference on Image Processing, 2008.
	A. Martin and J-J Fuchs and C. Guillemot and D. Thoreau Sparse Representation for Image Prediction EUSIPCO, 15th European Signal Processing Conference, Poznan, Poland, September, 2007.

Webmaster:
Last time modified: 2008-05-27