Color Attention Clue Based Sensitive Information Detection In Images

With the rapid development of Internet and mobile communication technologies, it is more convenient for people to get information. However, the information is usually mixed with a large amount of sensitive information, including some pornographic images and reactionary text etc.. The sensitive information has brought adverse effects to the healthy of teenagers and social harmony and stability. Thus, how to filter the sensitive information in the spread images is an important research topic. In this paper, we study the sensitive information detection methods from two aspects: pornographic image detection and image text recognition. The research is related to computer vision and machine learning fields. The research results can provide technical support to the filtering of sensitive images.The skin color detection is often used as a preprocessing step in the traditional pornographic image detection algorithms, which ignore the color distribution of the sensitive organs. Therefore, we use Gaussian mixture model to extract the color features of human sensitive organs, and use the histogram of gradient as their shape feature. Then, the two features are fused to train the sensitive organ detector by the latent support vector machine based deformable parts model. The test results of the proposed detector are better than the HOG feature based method. The proposed detector improves the detection ratio and greatly reduces the false alarm ratio compared with the traditional bag of word model.The existing methods of text location often neglect the color information contained in the image. Considering of the fact that characters with same color in the image tend to be distributed in a near position, we reduce the number of undetected characters based on this information. In this paper, the input image is first processed by the stroke width transform to obtain the candidate characters, then the color of each character is used as the a color prior of maximally stable extremal region to obtained the potential connected regions of characters. Compared with the traditional character location method, the proposed method has advantages on improving the recall, which provides comprehensive text candidate region as much as possible for the subsequent text recognition.The traditional character classifiers are not good at handle characters with large geometric deformation. With the sound development of deep learning methods, this problem has a better solution. In this paper, First, we get the connected regions which are based on the location, size and other relations. Then, we remove the candidate regions with low probability. Finally, we use the well trained deep learning model to get accurate text area location and recognition. We test the model on standard testing dataset to achieve good results. The recognition algorithm is combined with the results of the proposed text detection method to improve the accuracy of text localization.