基于人体关键点的色情信息检测

Pornographic Information Detection Based on Key Points of Human Body

传统的色情检测在色情图像领域取得了较高的准确率,但对于视频中未出现肤色裸露但含有性暗示动作的色情信息检测难以实现。为解决这个问题提出一种基于人体关键点的色情信息检测框架。首先采用人体姿态估计算法获取人体关键点,然后利用可变形卷积、肢体2D矢量场进行身体部位定位和关联程度分析,形成关键点序列、提取人体轮廓,最后通过RNN对二种附加的时空姿势特征以及人体关键点轨迹叠加融合,完成视频中色情信息的检测。实验结果表明,该方法在Pornography-2k数据集上达到81.7%的准确率,能有效地检测视频中并未出现大面积肤色裸露却含有性暗示动作的色情信息。

Traditional pornographic detection has achieved high accuracy in the field of pornographic images,but it is difficult to detect pornographic information that does not show naked skin but contains sexually suggestive actions.To solve this problem,a pornographic information detection framework based on key points of the human body is proposed.First,the human body pose estimation algorithm is used to obtain the key points of the human body,then the deformable convolution and the limb 2D vector field are used for body part positioning and correlation analysis to form the key point sequence and extract the human body contour,finally,the two additional spatio-temporal poses are calculated by RNN Features and key points of the human body are superimposed and integrated to complete the detection of pornographic information in the video.Experimental results show that this method achieves an accuracy of 81.7% on the Pornography-2k dataset,and can effectively detect pornographic information that does not appear in the video with a large area of naked skin but contains sexually suggestive actions.