Kernel Laplacian Sparse Coding For Image Classification

Image classification is a branch of the image processing field, its main research content is how to classify a large number of chaotic digital image automatically through the computer. Image understanding and image representation based on computer vision technology are the basis of classification. The image classification has been greatly improved since sparse coding theory is introduced into the image representation.Sparse coding is an effective and general coding scheme, the principle is to use the bases of over-complete dictionary to linearly represent image. For each code coefficient, the non-zero elements account for only small part of all elements, which represent the sparsity. The reason for sparse coding can replace k-means algorithm to generate the dictionary is that sparse coding algorithm in dictionary learning optimization process achieves an optimal weighting of the bases of dictionary. Hence, by doing this, both the quality of the dictionary and the quantification features accuracy are greatly improved. However, sparse coding algorithm is still not perfect, it ignores the relevance of features in the process of encoding. This may lead to similar features produce very different coding and then affect the classification results. Based on this, the researchers found that through adding a constraint condition into the sparse coding optimization function, which effectively make the similar features have similar sparse coding and reduce the quantization error of features. This can further improve the accuracy of image classification. This algorithm is an improvement for sparse coding and is called Laplace sparse coding by researchers.We find that the current sparse coding algorithm is used in the raw feature space. However, the kernel method can non-linearly map the features to high dimensions. Inspired by it, in our paper, we introduce the kernel method in sparse coding process and propose the kernel sparse representation. This method not only can reduce the quantization error of features, but also can enhance the performance of sparse coding. By combining Laplace sparse coding and the kernel sparse representation, we propose kernel Laplace sparse coding. Experiments show that our method is better than the Laplace sparse coding and have a good performance in images classification on the Caltech-101、Scene-15and UIUC-Sport datasets.