Fine-grained pornographic image recognition with multiple feature fusion transfer learning

Image has become a main medium of Internet information dissemination, makes it easy for an Internet visitor to get pornographic images with just few clicks on websites. It is necessary to build pornographic image recognition systems since uncontrolled spreading of adult content could be harm to the adolescents. Previous solutions for pornographic image recognition are usually based on hand-crafted features like human skin color. Hand-crafted feature based methods are straightforward to understand and use but limited in specific situations. In this paper, we propose a deep learning based approach with multiple feature fusion transfer learning strategy. Firstly, we obtain the training data from an open data set called NSFW with 120,000+ images. Images would be classified into different levels according to its content sensitivity. Then we employ data augment methods, train a deep convolutional neural network to extract image features and conduct the classification job, without the need for hand-crafted rules. A pre-trained model is used to initialize the network and help extract the basic features. Furthermore, we propose a fusion method that makes use of multiple transfer learning models in inference, to improve the accuracy on the test set. The experimental results prove that our method achieves high accuracy on the pornographic image recognition and inspection task.