基于图像增强与改进Cascade R-CNN的钢轨表面缺陷检测  

Rail Surface Defect Detection Based on Image Enhancement and Improved Cascade R-CNN

作　　者：罗晖[1] 李健[1] 贾晨 Luo Hui;Li Jian;Jia Chen(School of Information Engineering,East China JiaoTong Univevsity,Nanchang,Jiangxi 330013,China)

机构地区：[1]华东交通大学信息工程学院,江西南昌330013

出　　处：《激光与光电子学进展》2021年第22期316-327,共12页Laser & Optoelectronics Progress

基　　金：江西省重点研发计划项目(20202BBEL53001)。

摘　　要：针对钢轨表面缺陷检测中,钢轨表面图像存在背景不均匀、缺陷尺度变化大且样本数据不足的问题,提出一种基于图像增强与改进Cascade R-CNN的钢轨表面缺陷检测方法。首先,采用改进Retinex算法处理钢轨表面图像,增强缺陷与背景的对比度。然后,采用改进Cascade R-CNN对钢轨表面缺陷进行检测,并应用交并比(IoU)平衡采样、感兴趣区域对齐和完全交并比(CIoU)损失分别解决训练样本IoU分布与困难样本IoU分布不平衡、感兴趣区域池化中取整量化导致的感兴趣区域与提取的特征图不匹配和回归损失Smooth L1对于预测边框回归不准确的问题。最后,采用翻转变换、随机剪裁、亮度变换和生成对抗网络等方法增广钢轨表面缺陷图像数据集,消除样本数据不足导致的网络训练过拟合现象。实验结果表明,该方法以ResNet-50作为特征提取器,平均精度可达98.75%,相对于未改进的Cascade R-CNN提高了2.52%,且检测时间缩短了24.2 ms。In the rail surface defect detection,the rail surface image has the problem of uneven background,large variation of defect scale,and insufficient sample data.Therefore,this paper proposes a rail surface defect detection method based on image enhancement and improved Cascade R-CNN.First,the improved Retinex algorithm is used to process the rail surface image to enhance the contrast between the defects and the background.Then,an improved Cascade R-CNN is adopted to detect rail surface defects,and the intersection over union(IoU)balanced sampling,region of interest align and complete intersection over union(CIoU)loss are applied to solve the imbalance between training sample IoU distribution and the difficult sample IoU distribution,the misalignment between region of interest and extracted feature map caused by rounding quantization in region of interest pooling,and the inaccuracy of the regression loss Smooth L1 for the regression of predicted bounding box.Finally,the dataset of rail surface defect images is expanded using methods such as flipping transformation,random cropping,brightness transformation,and generative adversarial networks,so as to solve the phenomenon of over-fitting of network training caused by insufficient sample data.Experimental results show that the average accuracy of the proposed method,using ResNet-50 as the feature extractor,can reach 98.75%,which is 2.52%higher than the unimproved Cascade R-CNN,and the detection time is reduced by 24.2 ms.

关 键 词：图像处理 钢轨表面缺陷 图像增强 CascadeR-CNN 感兴趣区域对齐 完全交并比 数据增广 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]