Research On Pornographic Image Recognition Based On Weak Detection Mechanism And Finegrained Features

With the rapid development of the Internet and the growth of the number of smartphone users,it is very easy to access and spread sensitive(pornographic)pictures through the Internet.Identifying and filtering sensitive content on the Internet is crucial for the healthy growth of minors and the development of the Internet.In recent years,sensitive image recognition methods based on convolutional neural networks have basically replaced traditional methods.Such methods usually treat sensitive image recognition as a general image classification task,which is to extract features from the entire image and classify them.However,due to the particularity of the sensitive image,the sensitive parts generally appear in a part of the entire image,so only relying on the features extracted from the global will make the recognition accuracy rate too low.To solve this problem,a convolutional neural network model based on weak detection mechanism is first proposed.The model is based on the local detection mechanism of the object detection algorithm and the model training strategy based on multi-instance learning.The model can perform global and local image recognition.By using the weak detection loss function and the corresponding training strategy,the model can be trained only by using image-level labels,without the need to mark the object boxes.Then the problem is considered from the perspective of fine-grained classification,and a sensitive image recognition algorithm based on fine-grained features is proposed.This algorithm can gradually tap the fine-grained features of key parts in the image from coarse to fine,thereby improving the performance of sensitive image recognition.In order to achieve fine-grained feature extraction,a main body location module based on attention mechanism and a key part selection module based on sliding window are proposed.By combining the classification results of global features,object features and key part features,the finegrained model can make more accurate identification.In order to verify the effectiveness of the weak detection model,a large number of experiments were performed on the WDPorn and Poster Porn datasets.The experimental results show that the weak detection mechanism can significantly improve the recall rate of the model on sensitive images and the accuracy of overall recognition,and experiments on public dataset Pornography-800 compared with some existing advanced methods show that the weak detection model surpasses most of the current advanced algorithms.In order to verify the effectiveness of the fine-grained model,a large number of experiments were conducted on the Tiny Nude dataset and Pornography-800,which confirmed the promotion of fine-grained features on the accuracy of sensitive image recognition.The results of the experiment fully proved the effectiveness of the fine-grained module.