Research On Some Key Technologies Of Erotic Image Recognition

Purification of the network environment, and the protection of adolescent health is not only a government issue, but also important technical problems. Identification and filtering technology of sensitive image has been a great deal of concern,and has become an important research topic. To this end, we make analysis and research on sensitive image recognition technology, and give a number of experimental results. These techniques include color detection, texture detection, face detection, a significant regional extraction, image feature extraction, classifier design.Skin color detection is key to sensitive image recognition, color detection has a direct bearing on the accuracy of the final image classification results. We turn first to the properties of color, color space, color model study and analysis of several commonly used color model, and put forward a detection method based on the color YCgCr color space. Experiments show that the method can detected skin color region effectively in a different light from the complex background .For the issue of false detection in color detection, based on first-order gray level statistics of the color texture detection skin color mask is processed to eliminate the false color of the region,and to improve the accuracy of the color detection.The purpose of face detection is to reduce the false rate with face contained images, thus enhancing the detection accuracy in addition to face regions. By analyzing Several commonly used methods of face detection,a combination of YCgCr space color information and AdaBoost face detection algorithm is put forward. The experiments show that the algorithm compared with the traditional AdaBoost-based face detection improve the accuracy,and effectively reduce the error rate.The choosing of features plays a vital role in the the classification. In the feature extraction stage, we first extract mask based on the color features, and a few key feature extraction methods are described, and then using a improved skin color information based on the voting mechanism to extract salient image regions, and color, texture feature are extracted from the salient area. These features are used as a feature vector, and input to the Support Vector Machines classifier. thus complete the identification of sensitive images.The selection and design of Classifier is the basic problems in pattern recognition, classifier ability directly determine the performance of sensitive image recognition. In choosing classifier, we selected SVMs classifier which have great advantages to address the small sample, nonlinear and high-dimensional pattern recognition and classification, the basic framework of sensitive image recognition algorithm based on SVMs is constructed, In the same experimental environment four experiments by different methods is done to determine the final method we used. The experimental results show that the sensitive image recognition algorithm we used have 89.38% accuracy rate; 90.25% precision rate; 11.02% false-positive rate, 3.0f/s recognition capability.