基于比例特征的网络不良图像过滤算法研究

A network pornographic image filtering algorithm based on ratio features

目前对网络不良图像进行过滤的方法中,裸露肤色比依然是常被利用的数据,但是其存在一些缺陷,如伪肤色的干扰、非关键区域肤色的干扰、整体裸露肤色比有时候不与裸露程度呈正相关等。本算法利用人体肤色的一致性和人体结构特点,计算了以下三个比例特征作为判决依据——人脸尺寸占比、用人脸肤色数据改进肤色模型后的整体裸露肤色比及人体关键区域的裸露肤色比,有效地解决了上述问题;同时,对人脸检测增加了合适的旋转图像的处理过程,使其对于倾斜很大的人脸也能较好地检测。经过测试,算法的识别率达到91.5%,基本能满足实际应用的需求。

Currently the exposed skin ratio is still frequently used by most network pornographic image filtering methods.However,it has some defects such as the interference of analogous skin color,the interference of non-key-area skin and sometimes the irrelevance of the whole exposed skin ratio and the exposed extent.We propose a network pornographic image filtering algorithm based on the consistence of skin color and human body structure,which calculates the face size ratio,the whole exposed skin ratio after the skin color is improved by face skin color data in the skin color model and the key-area exposed skin ratio.The problems mentioned above have been effectively solved by using these three ratio features.Besides,we add a rotation process to the face detection module,so it can detect severely inclined face effectively.Experimental results show that the recognition rate of the proposed algorithm can achieve 91.5%,which meets the practical demand.