A Deep Network for Pornographic Image Recognition Based on Feature Visualization Analysis

The study of pornographic image recognition plays an important role in protecting physical and mental health of young people. However, features extracted by traditional methods are so single that these methods are difficult to deal with the diversity of pornographic contents. Therefore, we propose a deep convolution neural network (CNN) for pornographic image recognition based on feature visualization analysis. In this paper, firstly we establish a pornographic database containing 80000 images. Then, we train a basic CNN model that can accurately recognize pornographic images. To face a variety of challenging scenes, we utilize a method based on a deconvolution network to visualize CNN extractive features. By using this visualization method, we make a system analysis of our recognition model performance, and optimize the model to make it more robust and accurate. Our experimental results suggest that the proposed method significantly outperforms the state-of-art methods.