Research On Weakly Supervised Spatial Pyramid Based Image Classification

Image Classification is a fundamental problem in computer vision, aims at classify images based on trained models. Based on feature they used, image classification algorithms can be roughly divided into global feature based image classification and local invariant feature based image classification. In this paper, we mainly focus on local feature based image classification; the basic framework we use is spatial pyramid model.Traditional spatial pyramid model has two problems:First, spatial pyramid formulation process is fully unsupervised, totally ignores lots of supervision exist in image classification. Yet these supervision are weakly supervision on bins level, cannot directly be used in image classification. Second, we can see from spatial pyramid formulation process that, we given equal weights to different bins,but weights of different bins are different.We proposed weakly supervised spatial pyramid model to solve the problems exist in traditional spatial pyramid match model. Different from previous spatial pyramid model, we don’t preserve all bins. We measure the importance of bins on some level by K nearest neighbor entropy, then we remove those bins whose k nearest neighbor entropy bigger than threshold. Analysis shows we get a spatial pyramid model with different weights for different bins. We prove that result kernel of our algorithm can preserve semi-definiteness, and we furture prove that its extension importance sampling on spatial pyramid can also preserve semi-definiteness. We test our algorithm on two common used datasets. Result show that our algorithm is better than traditional spatial pyramid model...