基于改进DeepLabV3+的铁谱图像语义分割  

Semantic Segmentation of Ferrograph Images Based on Improved DeepLabV3+

作　　者：程亮 王静秋[1] 刘信良 孙康[1] CHENG Liang;WANG Jingqiu;LIU Xinliang;SUN Kang(College of Mechanical and Electrical Engineering,Nanjing University of Aeronautics and Astronautics,Nanjing 210016,China)

机构地区：[1]南京航空航天大学机电学院,江苏南京210016

出　　处：《机械制造与自动化》2021年第6期103-105,122,共4页Machine Building & Automation

摘　　要：针对传统铁谱图像分割方法需要人工设计特征、存在特征冗余及误差累积等缺点,提出一种改进的DeepLabV3+方法,实现了对铁谱图像中严重滑动、疲劳、切削等6种磨粒的语义分割。该方法在DeepLabV3+基础上,引入密集连接的空洞空间池化金字塔模块,以增大感受野;通过改进密集金字塔模块,采用互质膨胀率的空洞卷积,避免了栅格效应;采用全连接条件随机场以改善分割结果。实验结果表明:该方法在铁谱图像测试集上的平均交占比为87.1%。To overcome the shortcomings of traditional ferrograph images segmentation methods such as artificial design features,feature redundancy and error accumulation,an improved DeepLabV3+method is proposed to realize the semantics of six kinds of wear particles like severe sliding,fatigue and cutting particles in ferrograph images segmentation.The proposed method based on DeepLabV3+introduces densely connected spatial pooling pyramid module to increase the receptive field.By improving the dense spatial pooling pyramid module,atrous convolution with co-prime dilation is adopted to avoid grid effect.Conditional random field fully connected is applied to improve the segmentation results.Experimental results show that the mean intersection over union of the improved method on the test dataset of ferrograph images is 87.1%.

关 键 词：铁谱图像 语义分割 密集空间池化金字塔 条件随机场 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]