Detect and Depornize Pornographic Images Using Pre-trained CNN Models

Abstract

The ubiquitous Internet has achieved great transparency of information; however, it also becomes the media of negative information, especially pornographic content. Classifying and filtering out such content are of high demand on the social networks. CNN models can be used to classify pornographic images with good flexibility and high accuracy. However, building CNN models from scratch is complex and computationally expensive. Fortunately, training pre-trained CNN models using the technique of transfer learning takes relatively less time to develop and can still deliver great performances. In this paper, we show that the most recent pretrained CNN models can be re-trained to be pornographic image detectors using the technique of transfer learning. Furthermore, considering the practical applications in social media and entertainment, this paper introduces a depornize algorithm to cover these sensitive and pornographic content. Experimental results demonstrate the efficiency of both algorithms proposed in this paper.