Color pornographic image detection based on color-saliency preserved mixture deformable part model

To utilize the rich semantic information of sexual organs, we propose a new framework for pornographic image detection based on sexual organ detectors. Traditional sexual organ detectors are built on shape features. Since the color distribution of sexual organ in same pose is consistent, color is an important visual clue to represent sexual organs. We use color attribute to describe the local color of sexual organs and concatenate it with histogram of oriented gradients based shape feature to represent sexual organs. Based on the concatenated feature, we train sexual organ detectors by the color-saliency preserved mixture deformable part model (CPMDPM). We detect pornographic images sequentially with sexual organ detectors. In experiments, the optimal part number of the deformable part model is chosen experimentally. We evaluate the performance of each CPMDPM based sexual organ detector, which is superior over the shape feature based detector. The proposed pornographic detection method is superior over methods based on low level features of skin regions, bag of words model and color incorporated SIFT features etc.