Pornographic Image Recognition in Compressed Domain Based on Multi-Cost Sensitive Decision Tree

Most pornographic image recognition researches focus on detection accuracy. However, as the highly increasing of web data, detection speed becomes a new consideration. In this paper, the new issue is discussed from the following two aspects: 1) feature extraction in compressed domain and 2) classifier design, and then a simple, novel and yet effective pornographic image recognition method in compressed domain is proposed, which is based on multi-cost sensitive decision tree. More specifically, some features, including: features based on skin color region, features based on the results of image retrieval, features based on face and regions of interesting as well as global texture and color features, are extracted from the compressed image firstly. Afterward, a multi-cost sensitive decision tree construction algorithm is presented, based on which the decision tree of pornographic image recognition is established. Experimental results show the proposed method can not only effectively improve the detection accuracy but also the detection speed.