Pornographic image identification method based on multi-step identification and fusion key portion detection

Abstract

The present invention provides a pornographic image identification method based on multi-step identification and fusion key portion detection, belonging to the design image processing and computer vision technology field. The method comprises: employing a network spider technology and a data augmentation method to construct a pornographic image database; employing a database to perform fine regulation of a GoogLeNet network, and obtaining an image cartoon / non-carton dichotomy model; employing a residual network to perform fine regulation on a selected training set, and obtaining the normal / pornographic dichotomy model of a carton-type image and the normal / pornographic dichotomy model of a non-cartoon image; and finally performing marking of a naked chest for the database image, employing a target detection network Faster RCNN to perform training and obtaining of a chest detection model, and performing secondary identification of the detected chest through cascading of a classification network after the Faster RCNN so as to ensure the accuracy of the chest detection and allow a pornographic image having the naked chest and a small skin color area not to be undetected.

Description

technical field [0001] The invention relates to a pornographic image recognition method based on step-by-step recognition and fusion key part detection, and belongs to the technical field of design image processing and computer vision. Background technique [0002] Common technologies in the field of image recognition can be divided into methods based on machine learning and methods based on deep learning. Machine learning methods mainly include feature extraction and classifier modules. The process of manually designing and adjusting features is cumbersome and relies on a large amount of engineering technology and professional knowledge. The quality of features directly affects the performance of the algorithm; deep learning methods combine feature extraction and classifiers. In one framework, the cumbersome steps of manually designing features and classifiers are no longer required. Its deep nonlinear network structure has strong feature expression capabilities and can lea...

AI Technical Summary

Problems solved by technology

This type of method has a high false alarm rate for normal images with large areas of bare skin and pseudo-skinned images, and the artificially designed pornographic image features are not effective enough, so the recognition effect is not ideal

At present, deep learning methods have been applied to the problem of pornographic image detection, which has achieved high recognition accuracy, but the existing pornographic image recognition methods are not suitable for the recognition of black and white comic pornographic images; and due to the deep convolutional neural network layer by layer The design of convolution and pooling (pooling) is prone to feature loss when the area of ​​sensitive parts in the image is small, resulting in missed detection of pornographic images

Images