Development of a real-time nudity censorship system on images

We propose a system for detecting and localizing nudity content in images and series of images (video), so that further action can be done on relevant segments (typically covering). We also attempt to make this system able to be run on video input, compromising frame rate as little as possible. In this paper we use a skin filtering method based on a Bayes rule, followed by a novel histogram backprojection of skin samples selected based on standard deviation and gradients. The filtered image is then clustered into a Bag-of-Visual-Words (BOVW) and classified using SVM. Localization is then achieved by shrinking the image window borders on which the classification is performed. The skin filtering method described here has an increased recall rate compared to filtering using pure Bayes rule. The nudity classification phase achieved a highest accuracy rate of 76.11% on a 5-fold cross validation test. Finally, the localization phase, when run on video clips, has successfully localized a minimum of a quarter of nude segments in a clip, although many cases of over segmentation still occur. The system is able to maintain a highest average frame rate of 15 FPS when run on videos using hardware specifications that we used.