基于改进ShuffleNetV2的敏感内容识别与应用

Sensitive content recognition and application based on improved ShuffleNetV2

针对目前公共场合大屏显示系统视频内容审核方法识别准确率低、难以部署在控制器上的问题,提出一种基于改进ShuffleNetV2的敏感内容识别方法。首先,在Block2模块拼接特征通道后引入高效通道注意力(ECA)模块,加强重要特征通道的权重;其次,采用最大池化替换Block2模块中的深度可分离卷积,减少复杂背景的干扰。将训练得到的模型进行转换并通过参数量化压缩模型,部署在以RK3399Pro为核心处理器的嵌入式控制器上,设计应用程序实现对视频文件中敏感内容的识别。实际测试结果表明:改进的ShuffleNetV2敏感内容识别模型准确率提升了3.85%,计算量减小了12.99%,在控制器上的检测速度达到每帧图像17 ms,并取得较好的识别效果,该方法可有效审核视频内容,并为大屏显示系统视频内容安全提供了可靠保障。

Aiming at the problem that recognition accuracy of video content audit method of large screen display system in public places is low and it is difficult to deploy on controller, a sensitive content recognition method based on improved ShuffleNetV2 is proposed.Firstly, efficient channel attention(ECA) block is introduced after the feature channels of Block2 are spliced to strengthen the weights of important feature channels. Secondly, the deep separable convolution of Block2 is replaced by maximum pooling to reduce the interference of complex background.The trained model is converted and compressed through parameter quantization and deployed on the embedded controller using RK3399Pro as the core processor.The application program is designed to realize the recognition of sensitive content in the video file. The actual test results show that the accuracy of the improved ShuffleNetV2 sensitive content recognition model is increased by 3.85 %, and the amount of calculation is reduced by 12.99 %. The detection speed on the controller reaches 17 ms per frame of image and achieves good recognition effect. This method can effectively audit the video content and provide a reliable guarantee for the safety of the video content of the large screen display system.