La classification dans les sous espaces pour l’analyse d’images

Abstract

Clustering problem consists in partitioning a given data set into groups called clus¬ ters, such that the data points in a cluster are more similar to each other than points in different clusters. The clustering in high-dimensional data is extremely difficult. In high dimensional datasets, the clusters can be characterized only by some dimensions subsets. These relevant dimensions can be different from one cluster to another. A new challenging research field has emerged, namely the subspace clustering. It is an extension of traditio¬ nal clustering that seeks to find clusters in different subspaces within a dataset. Image processing and image analysis tools are widely used in different domains. However, exploi¬ ting these images is tightly dependant of their textures. In this work, we have developed two approaches to image classification. The first one is a subspace clustering method. It is an iterative algorithm based on the minimization of an objective function. This function is formed by a separation and compactness terms. The cluster density is also introduced in the compactness term. An initialization step has been improved by a multi class SVM algorithm. An active learning with SVM is incorporated in the classification process to speed the proposed algorithm convergence. It allows enhancing the cluster center loca¬ tion. The second approach is based on a new non linear model which extends the random coefficients autoregressive model (RCA) to a bidimensionally RCA model (2D-RCA).The coefficients are estimated by the generalized moments method (GMM). It is a supervised method. We have proposed different versions of classification algorithms. The developed approaches have been tested and evaluated on different synthetic datasets and textures and real images. Experimental results have corroborated the effectiveness of the proposed method compared to well-established and state-of-the-art methods.