Learning Hierarchical Features Using Sparse Self-organizing Map Coding for Image Classification

Image descriptor is a critical issue for most image classification problems. Low-level image descriptors based on Gabor filters, SIFT and HOG features have exhibited good image representation for many applications. However, these descriptors are not appropriate for image classification and need to be converted into appropriate representations. This process can be performed by applying two operations of coding and pooling. In coding operation, an appropriate codebook is learned and better adapted from training data, while pooling operation summarizes the coded features over larger regions. Several methods of coding and pooling schemes have been proposed in the literature. In this paper, self-organizing map (SOM) is employed to learn a topologically adapted codebook instead of the well-known k-means algorithm. In addition, a new non-negative sparse coding technique for SOM is also proposed and tested. The new sparse coding method utilizes Non-negative Least Squares (NNLS) optimization to best reconstruct every input pattern using K-best-matching codewords. Experimental results using Caltech-101 database show the effectiveness of the proposed method compared with other state-of-the-art methods.