A New Image Retrieval Algorithm Based on Sparse Coding

摘要

The Bag-of-visual-words (BOVW) model discards image spatial information, and the computing cost is expensive on spatial pyramid matching(SPM) model. Due to sparse coding approach exhibit super performance in information retrieval, hence, we propose a new sparse coding image retrieval algorithm. Using l_2 norm replace l_0 norm in SPM vector quantization. The local information was incorporated into sparse term by local adapter. Sparse coding was transformed into least square convex optimization proplem. Each block was encoded by the k nearest neighbor (KNN) approach, and the coding coefficients were integrated by the max pooling function. Each block required different weight according to the image itself information. Euclidean distance and the cosine theorem were combined with the similarity calculation. Our method is evaluated on the two datasets - Caltech-101 and Corel-1000. Comparing with the BOVW and SPM, the results are shown that the new approach greatly improves the image retrieval accuracy.