Pornographic image classification based on top down color-saliency based BoW representation

Since color is an important visual clue of the pornographic image, this study presents a new framework for pornographic image classification based on the fusion of color and shape information for the bag of words representation. This framework contains three fusion patterns: The early fusion, late fusion and top down color-saliency based fusion, which are compared intensively. Based on the comparison, the top down color-saliency fusion based pornographic image classification method is proposed by using the statistical class prior of each color word to weight the shape word. In the late fusion and color-saliency based fusion, color name is adopt to represent the color information. To verify the effectiveness of spatial constrain on the words, we also compared the shape features quantized by vector quantization and locality-constrained linear coding. The experimental results show that our model combines the shape and color information properly and it is superior over the popular methods to distinguish the normal and pornographic-like images from the pornographic ones.