Fast and Effectively Identify Pornographic Images
Yanjun Fu, Weiqiang Wang · 2011
In this paper, we present a practical solution to identifying pornographic images based on multiple low-level image features and support vector machine (SVM). First, the region of interest (ROI) is obtained from an original image based on the detection of skin-like pixels in YCbCr color space. The ROI is then classified into being acceptable or unacceptable by its size. Images without ROIs or acceptable ROIs are deemed to be benign images. For images with acceptable ROI, the color, texture and shape features are further extracted on ROI, and then fed to a SVM classifier to perform the task of recognition. Our approach has obtained 96.05% sensitivity and 96.17% specificity on a dataset containing 8,000 pornographic images and 12,500 benign images, as well as the processing speed of about 0.026 seconds for a PC to determine whether a given image with an average size of 420 by 433 pixels is pornographic.