Image Classification Based on Region Non-uniform Spatial Sampling

Extensive experiments demonstrate that locally dense features are able to improve greatly performances of image classification, and the popular way is to conduct spatially uniform sampling for locally dense feature extraction. In this paper, a new method to extract locally dense features, region-based non-uniform spatial sampling is proposed to improve further the performance of image classification. Firstly, an over-segmentation operator is performed on the image, and then a saliency detection method is applied to estimate the importance of each segmented region. To keep the same sampling number of local features, the dense features are extracted along the boundary of the important salient region with dense sampling, as well as inside the region with random sampling according to its area and importance. Finally, the Bog-of-Words representation model is used for image classification. Extensive experiments are conducted on two widely-used datasets (UIUC Sports and Caltech-256). The experimental results show that proposed sampling strategy obtains an efficient performance.