Réseaux de neurones dans la statistique exploratoire et à l’aide de la décision.

Abstract

Data analysis (also called exploratory data analysis) is a family of statistical methods whose main characteristics are that they are multidimensional and descriptive. These methods can also be considered as special neural methods. In this thesis work, a focus on the statistical aspects of neuronal methods is proposed. As an innovative contribution in this field, a region growing technique is used to achieve image segmentation by merging some starting points or internal small areas if they are homogeneous according to a measurement of a local region property. A 2D random coefficients autoregressive model (2D RCA) is fitted in order to identify the different textures present in the image. First, an estimation procedure using a generalized method of moments (GMM) technique is proposed to extract some local region properties. For this, a gradient-based neural network (GNN) is used to estimate the 2D RCA model parameters from a given texture. The cost function of the proposed (GNN) is based on a strong matching of the statistical moments of the corresponding 2D-RCA model and the sample moments of population image data. Experimental results demonstrate the effectiveness and the relevance of the proposed method.