The Method Of Integrating Local Sensitive Feature Detection And Global Perception Classification To Detect Pornographic Images

With the Development of Internet,pornography images have a tendency to overflow.Traditional methods of identifying simple skin tones and texture features have failed to identify these pornographic images quickly and effectively.With the development of Deep Learning and Convolutional Neural Networks,there has been a new breakthrough in the direction of pornographic image recognition,the recognition efficiency,recognition accuracy,and recognition range are expanding.However,existing pornographic image detection methods either focus on the sensitive parts of the image or the overall semantic expression of the image,only for a certain type of image,and cannot verify and identify mass and irregular pornographic images on the Internet.Real-People pornographic images can be categorized into two types,including nudity with sensitive parts or nudity without sensitive parts,but the overall semantic expression is pornography,including sexual cues,sexual postures,etc.In order to better identify the various forms of live pornographic images spread on the Internet,consider the characteristics of different types of pornographic images.In this paper,by combining the characteristics of the local field and the characteristics of the global field,and through the second discriminant method,we can more accurately identify live pornographic images.The main research content and innovation of this paper include:(1)We propose a method using different fields of view is proposed to judge thenature of the image.By using local sensitive feature detection and globalperception classification,that is,combining the local features and overallsemantics of the image to determine whether the image is a pornographic image.(2)We propose a secondary discrimination strategy based on image classification.In view of the possible missed inspections of the model,that is,the model usestwo methods to judge all images as normal,and uses the probability average ofthe porn category output by the global classification module to make a secondjudgment to improve the overall recall rate and reduce missed inspectionExperiments show that this method not only improves the identification accuracy of pornographic images with obvious sexual organ bareness,but also has a good recognition effect for all forms of real-life pornographic images without sexual organ bareness but with strong sexual implications.On 200,000 data sets randomly collected on the network,the recall rate and precision rate of the model are 93.7% and 87% respectively.On the public pornographic dataset NPDI,the recall rate and the precision rate are 87.6%,and 90.3% respectively.