In recent literature of image analysis

it has been very popular to code local features into visual words. We propose a novel kernel which fuses multi-level contexts of visual words. Besides the histogram pyramid of words

our kernel also incorporates the histogram pyramid of visual phrases （the local co-occurrence patterns of words） and the context classes of those words and phrases. Then support vector machines using the kernel are trained to perform image classification. Our method performs well on a wide range of test data

such as the Corel dataset. The method is also tested in a challenging problem

the discrimination of pornographic images from bikini ones. The classification accuracy of our method is 7% higher than that of the baseline method. Experimental results demonstrate that the performance of image classification can be improved by the integration of kernel based measurements and the multi-level representation of visual words. In the future work

more compact and efficient representation of contexts should be researched.