Adult Image Classification Based On Deep Learning

Internet technology has surged in the 21 st century,and various resources have been widespread on the Internet.While the Internet brings convenience to the acquisition and dissemination of resources,it also incurs many hidden dangers.The spread of sensitive images is one of the biggest hidden dangers,and its impact on adolescents is specifically harmful.Therefore,in order to prevent young people from accessing sensitive images,automatically identifying and filtering sensitive images has become an increasingly important issue and an important part of purifying cyberspace.However,due to the variety and complex nature of sensitive images,sensitive image recognition and analysis task becomes even more challenging.In this paper,a sensitive image recognition method based on deep convolutional neural network is proposed,which classifies images into three categories,namely normal,sensitive and sexy.The proposed method takes into account both the global information of the image and the sensitive body region information of the image.The designed network architecture consists of four parts,which are a global classification network based on ResNet and the feature pyramid network used to extract the global features of the image,a sensitive body part detection network to detect the local information related to the sensitive body region,an attention mechanism network to extract high-level discriminative feature from the original image,and a feature extraction and fusion network to generate high-level and discriminative feature for classification.Furthermore,a multi-task learning mechanism is applied to optimize the global classification network and local sensitive body part detection network simultaneously.The curriculum learning mechanism is used to improve the generalization ability of the designed network architecture.In this paper,experiments were conducted in the public data set NPDI and the self-acquired data set AIC,for a total of more than 160,000 images.The experimental results show that the proposed network has a classification accuracy of 96.6% on AIC and 92.7% on NPDI.This result exceeds other sensitive image recognition methods and verifies the effectiveness of the proposed method.