基于DeepLabV3+的轮对踏面损伤分割算法  

Image segmentation of wheel set tread damage based on DeepLabV3+

作　　者：孙耀泽 高军伟[1,2] Sun Yaoze;Gao Junwei(School of Automation,Qingdao University,Qingdao 266071,China;The Shandong Province Key Laboratory of Industrial Control Technology,Qingdao 266071,China)

机构地区：[1]青岛大学自动化学院,青岛266071 [2]山东省工业控制技术重点实验室,青岛266071

出　　处：《电子测量技术》2022年第23期113-118,共6页Electronic Measurement Technology

基　　金：山东省自然科学基金(ZR2019MF063);山东省重点研发计划(2017GGX10115)项目资助。

摘　　要：针对轨道交通轮对踏面损伤图像存在边界识别效果差,分割精度低的问题,提出一种改进的DeepLabV3+算法对损伤区域进行识别分割。首先将轻量化网络MobileNetV2作为主干特征提取网络,加快语义分割的速度;然后将空洞空间卷积池化金字塔模块中的膨胀卷积以及特征融合后的普通卷积替换为深度可分离卷积,减少参数量,降低模型复杂度;最后在主干网络输出的浅层与深层特征层添加ECA机制,加强网络特征学习能力,使模型对损伤区域更加敏感,从而提升模型分割精度。实验结果表明,改进后的DeepLabV3+模型大小缩减为原来的5%,mPA值达到90.70%,mIou值达到84.33%,在模型更轻量化的同时保证了踏面损伤图像的分割效果。Aiming at the problems of poor boundary recognition effect and low segmentation accuracy of rail transit wheel set tread damage image, an improved DeeplabV3+ algorithm is proposed to recognize and segment the damage area. Firstly, the lightweight network MobileNetV2 is used as the backbone feature extraction network to speed up the speed of semantic segmentation;Then, the expansion convolution in Atrous Spatial Pyramid Pooling module and the ordinary convolution after feature fusion are replaced by Deep Separable Convolution, so as to reduce the amount of parameters and reduce the complexity of the model;Finally, ECA mechanism is added to the shallow and deep feature layers of the backbone network output to strengthen the network feature learning ability and make the model more sensitive to the damaged area, so as to improve the segmentation accuracy of the model. The experimental results show that the size of the improved DeeplabV3+ model is reduced to 5%, the mPA value is 90.70%, and the mIou value is 84.33%. While the model is lighter, the segmentation effect of tread damage image is ensured.

关 键 词：轮对踏面损伤 DeeplabV3+算法 MobileNetV2 ECA注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]