Spatial locality-preserving feature coding for image classification

摘要

The state-of-the-art image classification models, generally including feature coding and pooling, have been widely adopted to generate discriminative and robust image representations. However, the coding schemes available in these models only preserve salient features which results in information loss in the process of generating final image representations. To address this issue, we propose a novel spatial locality-preserving feature coding strategy which selects representative codebook atoms based on their density distribution to retain the structure of features more completely and make representations more descriptive. In the codebook learning stage, we propose an effective approximated K-means with cluster closures to initialize the codebook and independently adjust the center of each cluster of the dense regions. Afterwards, in the coding stage, we first define the concept of “density” to describe the spatial relationship among the code atoms and the features. Then, the responses of local features are adaptively encoded. Finally, in the pooling stage, a locality-preserving pooling strategy is utilized to aggregate the encoded response vectors into a statistical vector for representing the whole image or all the regions of interest. We carry out image classification experiments on three commonly used benchmark datasets including 15-Scene, Caltech-101, and Caltech-256. The experimental results demonstrate that, comparing with the state-of-the-art Bag-of-Words (BoW) based methods, our approach achieves the best classification accuracy on these benchmarked datasets.