摘要
面向复杂多样的钢轨场景,本文扩展了最先进的深度学习语义分割框架DeepLab v3+到一个新的轻量级、可伸缩性的贝叶斯版本DeeperLab,实现表面缺陷的概率分割.具体地,Dropout被融入改进的Xception网络,使得从后验分布中生成蒙特卡罗样本;其次,提出多尺度多速率的空洞空间金字塔池化(Atrous spatial pyramid pooling,ASPP)模块,提取任意分辨率下的密集特征图谱;更简单有效的解码器细化目标的边界,计算Softmax概率的均值和方差作为分割预测和不确定性.为解决类别不平衡问题,基于在线前景-背景挖掘思想,提出损失注意力网络(Loss attention network,LAN)定位缺陷以计算惩罚系数,从而补偿和抑制DeeperLab的前景与背景损失,实现辅助监督训练.实验结果表明本文算法具有91.46%分割精度和0.18 s/帧的运行速度,相比其他方法更加快速鲁棒.
This paper extends the state-of-the-art deep learning framework DeepLab v3+to a light-weighted and scalable Bayesian version DeeperLab for the defect detection on complex and diverse rail surface.Specifically,Dropout is incorporated into the improved Xception network for Monte Carlo sampling from posterior distribution.Atrous spatial pyramid pooling(ASPP)module is utilized to extract the dense features at multiple scales and rates.Furthermore,a simpler and efficient decoder is proposed to improve the defect edges,and outputs the mean and variance of Softmax probability as segmentation and uncertainty.To solve class imbalance problem,we present the loss attention network(LAN)to perform auxiliary supervision for DeeperLab training.Experimental results show that the proposed algorithm is more accurate and robust than other methods with 91.46%precision and 0.18 s/frame speed.
作者
金侠挺
王耀南
张辉
刘理
钟杭
贺振东
JIN Xia-Ting;WANG Yao-Nan;ZHANG Hui;LIU Li;ZHONG Hang;HE Zhen-Dong(College of Electrical and Information Engineering,Hunan University,Changsha 410082;National Engineering Laboratory of Robot Vision Perception and Control Technology,Hunan University,Changsha 410082;College of Electrical and Information Engineering,Changsha University of Science and Technology,Changsha 410114;College of Electrical and Information Engineering,Zhengzhou University of Light Industry,Zhengzhou 450000)
出处
《自动化学报》
EI
CSCD
北大核心
2019年第12期2312-2327,共16页
Acta Automatica Sinica
基金
国家自然科学基金(61573134,61733004)
湖南省科技计划项目(2017XK2102,2018GK2022,2018JJ3079)资助~~
关键词
钢轨表面缺陷
视觉检测
贝叶斯卷积神经网络
注意力机制
类别不平衡
Rail surface defects
visual detection
Bayesian convolutional neural network(CNN)
attention mechanism
class imbalance