Abstract: Non-pornographic images containing large naked skin or skin-like areas may be detected as pornographic using ordinary pornographic image-filtering method that heavily depend on skin detection. We design a different pornographic image-filtering system based on body parts. The system extracts Haar-like features describing local grayscale distribution of the body parts, and uses these features to train and obtain a classifier for body parts using the Adaboost learning algorithm. The candidate body part areas can be obtained with the classifier. To further improve the system performance, we extract histogram of oriented gradient features,textual features based on gray level co-occurrence matrix, and color moment features of the candidates. These
features are then used to train an support vector machine (SVM) classifier. Experiments show that the system can precisely detect key body parts in images, and therefore can effectively reduce false detection rate against non-pornographic images.

Key words: Haar-like feature, Adaboost learning algorithm, histogram of oriented gradient, gray level cooccurrence matrix, color moment