Research On Content-based Web Pornographic Image Recognition

With the continuous development of multimedia and network technology, a large amount of information can be browsed and shared freely on the Internet. Meanwhile, the harmful information containing violence, pornography, reactionary, especially pornography/video(Hereinafter referred to as pornographic image/video) has greatly harmed the social stability and people's physical and mental health, especially affected the healthy growth of teenagers. Considering the huge information of network, it has become the important research content to identify and filter the bad information automatically in the field of network information security.In order to curb the spread of pornographic image effectively, researchers have launched research on it and a variety of methods to identify or block pornographic image have been proposed, and those based on image content have become the mainstream image recognition methods. They embark from the image itself to analysis the content and features of pornographic image using digital image processing, pattern recognition, machine learning and etc. Automatic identification will be implemented by judging these difference features from other normal images. They can achieve good recognition performance. At present, these kind methods can be divided into four categories, where recognition method based on classification performance can achieve the best performance. This category is to consider the images contain two kinds of images(pornographic or non-pornographic images), and includes three parts:feature extraction, feature representation and the classifier. This paper has launched research on these three core parts respectively. The major contributions of this paper are as follows:(1)Oriented FAST and Rotated BRIEF(ORB) feature based web pornographic image recognition has been proposed. The whole recognition process can be divided into two parts: coarse detection and fine detection. The part of coarse detection contains skin color detection and face detection, and it can identify the non-pornographic images with no or fewer skin-color regions and facial images(such as ID photos) quickly. For the remaining images containing much more skin-color regions, fine detection is conducted, which includes three steps: first of all, extract ORB descriptors from the skin-color regions and extract 72-dimensional Hue, Saturation, Value(HSV) color feature from the whole image. And then, extracted ORB descriptors can be represented based on Bag of Words(BOW) model, and construct the feature vector combining with HSV histogram by linear weighted fusion method. Finally, train the classification model using Support Vector Machine(SVM) and apply it for image recognition. The experiment database contains 19000 images. Compared with Scale Invariant Feature Transform(SIFT), Speeded Up Robust Features(SURF) and so on five local features based on the same image recognition scheme, ORB can achieve a good trade-off between recognition speed and precision.(2)A pornographic image recognition method based on sparse representation has been proposed. The sparse coding method has been applied into the field of pornographic image recognition to replace the Bo W model for the feature representation. Experiments show that, compared with Bo W model method, this method can slightly improve the recognition precision, but time-consuming of recognition each image is huge.(3)In view of the successful application of the sparse classifiers in face recognition and other fields, the sparse classifiers are applied to the field of pornographic image recognition and a pornographic image recognition method based on sparse classifiers is proposed in this paper. In addition, In order to more accurate image description,, this paper also extracts the three global feature descriptor defined in MPEG-7- Color Structure Descriptor(CSD), Homogeneous Texture Descriptor(HTD) and Edge Histogram Descriptor(EHD) in addition to the ORB local feature and HSV color histogram, and apply them to represent the image together. The experimental results show that: Collaborative Representation based Classification with Regularized Least Square(CRC_RLS) can obtain relatively optimal recognition performance. It can achieve the highest average recognition precision of 96.38% for all images and the average 1015 ms to identify an image.(4)Combining with the study results of this paper, a pornographic image recognition demonstration system is set up, and the feasibility and effectiveness of the algorithm can be verified through this system in this paper.