Low-rank, sparse matrix decomposition and group sparse coding for image classification

This paper presents a novel image classification framework (referred to as LR-GSC) by leveraging the low-rank, sparse matrix decomposition and group sparse coding. First, motivated by the observation that local features (such as SIFT) extracted from neighboring patches in an image usually contain correlated (or common) items and specific (or noisy) items, we decompose the local features matrix of an image into a low-rank matrix and a sparse matrix. Second, we train the group sparse dictionaries on the low-rank parts and sparse parts respectively. And then, the dictionaries of the two parts are jointed to encode the original SIFT features by group coding. Finally, linear SVM classifier is used for the classification. The method is tested on the Caltech-101 dataset and UIUC-sports dataset, and achieves competitive or better results than the state-of-the-art methods.

1. INTRODUCTION

Image classification is one classical problem in computer vision. Among many methods, Bag-of-Words (BoW) [1] has been widely used by many researchers and shown good performance, which represents an image as a histogram of its local descriptors. It is robust against spatial translations of features. However, the nature of histogram discards all information about the spatial layout of local descriptors, which limits the representation power. An extension of bag-of-feature representation called spatial pyramid matching (SPM) [2] has received widely application, which partitions an image into several finer spatial sub-regions in different scales and computes the word histogram inside each of the sub-regions, and then concatenates all the histograms to form a vector representation of the image.