一种基于密集连接与感受野的穿戴识别深度学习网络

An Apparel Recognition Deep Learning Network Based on Dense Connection and Receptive Field

为实现在复杂环境中对电力作业人员穿戴规范的识别,以感受野块网络(Receptive Field Block Network,RFBnet)为基础网络设计了一种穿戴多目标识别网络。该网络引入特征金字塔网络并采用VoVnet的密接方式改进RFBnet的VGG-16主干网络,在保证浅层细节高分辨率的前提下尽可能获取深层语义信息以实现对弱小目标的检测。采用空洞卷积分支构建新的RFB模块以扩大感受野、提取更多的细节特征;将Soft-nms与DIoU-nms结合,设计一种后处理方法进行锚框回归,降低了漏检几率。实验结果表明,在RFBnet-512的基础上,该检测网络的mAP提升了10.09%,检测速度提升了4.7帧/秒;对安全帽、绝缘手套和高压绝缘靴的检测精度分别达到90.9%、67.1%和86.4%,检测速度达到17.5帧/秒。

In order to recognize the wearing of electric power industry workers in complex environment,a multi-target recognition network for wearing recognition based on Receptive Field Block Network(RFBnet)is designed.The proposed network uses FPN and dense connections method of VoVnet to improve the VGG-16 backbone network of RFBnet,thus obtaining deep semantic information as much as possible to detect small targets on the basis of ensuring high resolution of details in the shallow feature map.Using dilated convolution,a new RFB module is constructed to expand the receptive field and extract more detailed features.Combining Soft-nms with DIoU-nms,a post-processing method is designed to perform anchor frame regression,which reduces the undetected rate.Experimental results show that based on RFBnet-512,the mAP of the detection network is increased by 10.09%,and the detection speed is increased by 4.7 frame/s.The detection accuracy of safety helmets,insulated gloves and high-voltage insulated boots,respectively,reaches 90.9%,67.1%and 86.4%,and the detection speed reaches 17.5 frame/s.