Abstract An efficient image categorization method based on improved distribution of local invariant features is proposed. Firstly, the distribution of local features of an image is modeled by Gaussian mixed models (GMMs). Then, initial probability signatures are established by projecting local features onto all single models of GMMs. Finally, the complete probability signatures are obtained by performing a compression process. The probability signatures retain high discriminative power of probability density function (PDF) model, and they are suited for measuring dissimilarity of images with earth mover's distance (EMD), which allows for partial matches of compared distributions. The images are classified by learning support vector machine classifier with EMD kernel. The proposed method is evaluated on three image databases in scene recognition and image categorization tasks. The results of comparative experiments show that the proposed algorithm has inspiring performance.
2011, Vol. 24 
