Sparse coding based spatial pyramid matching algorithm for image classification

Image classification is a hot issue in computer vision field and has been extremely popular in mobile internet applications. A sparse coding based spatial pyramid matching(SCSPM) algorithm was proposed. Firstly, the extracted scaleinvariant feature transform(SIFT) descriptors are encoded by sparse coding method instead of the traditional vector quantization method. The sparse coding step can reduce the quantization errors effectively and generate more discriminative image representation. Furthermore, image classification can be obtained by linear spatial pyramid matching method. The experimental results on Caltech 101, Caltech 256 and 15 Scenes data sets show that our method can significantly improve the image classification accuracy by 4%~12% compared with the bagoffeatwes(BOF) and SPM algorithms.