Soft sparse coding neural response for image feature extraction

Image feature extraction is a key technique in image recognition and computer vision. In this paper, a soft sparse coding neural response (SSCNR) was proposed for image feature extraction. The main idea of the proposed method was to extract feature of the input image by alternation between soft sparse coding and maximum pooling operation within a hierarchical structure. Firstly, the feature of every image patch was extracted by soft sparse coding of the similarities between the patch and the templates and led to a sparse similarity matrix. Then, the maximum pooling operation carried on the similarity matrix and the transformation invariance was introduced in the model. In addition, a refined template selection method was proposed to reduce the computation complexity and to improve the recognition ability. Experimental results on several image databases strongly demonstrated the effectiveness of the proposed method.