基于改进DSD-LinkNet的电力安全带分割算法  

Power Safety Belt Segmentation Algorithm Based on Improved DSD-LinkNet

作　　者：马青山 朱建宝 俞鑫春 张斌 Ma Qingshan;Zhu Janbao;Yu Xinchun;Zang Bin(Nantong Power Supply Branch,State Grid Jiangsu Electric Power Co.,Ltd.,Nantong Jiangsu 226006,China;Nantong Huayuan Technology Development Co.,Ltd.,Nantong Jiangsu 226007,China)

机构地区：[1]国网江苏省电力有限公司南通供电分公司,江苏南通226006 [2]南通华远科技发展有限公司,江苏南通226007

出　　处：《电气自动化》2023年第3期106-108,共3页Electrical Automation

基　　金：江苏省电力公司南通供电公司科技项目“基于计算机视觉的电力作业现场违章行为识别技术研究”(J2020054)。

摘　　要：针对电力安全带分割检测背景干扰大、难以区分腰带和安全绳问题,提出了基于改进深度监督分割网络(deeply supervised D-LinkNet,DSD-LinkNet)的安全带分割算法。首先,为了解决深度网络训练消失和收敛速度过慢问题,在语义分割神经网络的基础上加入深度监督机制,为隐藏层引入伴随目标函数来提供梯度信息;其次,为减少图像像素点分布不均衡的影响,引入带权重的损失函数来平衡不同目标类别对损失函数值的影响。试验结果表明,改进后的网络能获得更精确的分割结果,更有实际应用价值。In order to solve the problem that the background interference was large and it was difficult to distinguish the waistband from the safety rope,a safety belt segmentation algorithm based on improved DSD-LinkNet was proposed.Firstly,in order to solve the problem of the disappearance of deep network training and the slow convergence speed,a deep supervision mechanism was added on the basis of D-LinkNet to provide gradient information for the hidden layers by introducing an adjoint objective function;secondly,in order to reduce the impact of uneven distribution of image pixels,a weighted loss function was introduced to balance the impact of different target categories on the loss function value.The experimental results show that the improved network can provide more accurate segmentation results with good practical application value.

关 键 词：安全带 深度学习 深度监督 损失函数 

分 类 号：TM71[电气工程—电力系统及自动化]