摘要
针对传统混凝土轨枕裂纹识别检测方法中存在效率低下与准确度差的问题,提出一种基于级联卷积神经网络的混凝土轨枕裂纹识别算法。该算法主要包含裂纹定位检测网络与裂纹显著性分割网络,其中裂纹显著性分割网络包含裂纹粗显著性分割模块CEDNet (Crack Encoder Decoder Network)和裂纹边界精修模块CRRNet (Crack Residual Refinement Network),实现对轨枕裂纹分割精确分割(轨枕裂纹的精确分割)。在有砟轨道道床图像中,混凝土轨枕裂纹边缘特征易淹没于碎石道砟边缘噪声中,因此,在输入SSD网络前先经过灰度投影法提取轨枕区域,然后用SSD网络对轨枕上的裂纹进行精确定位与分割。为进一步提取裂纹边缘特征的完整信息,对分割后的裂纹区域采用显著性分割算法,将其输入到CEDNet模块中先获得裂纹粗显著性预测图。将预测图输入到CRRNet模块,对其边缘信息与局部区域加以完善以达到保留边缘特征完整目的。采用混合BCE,SSIM和IOU 3种损失的裂纹检测综合损失函数,用来评价裂纹显著性分割网络提取出的裂纹与真实裂纹的偏差。实验结果表明:采用该综合损失函数,本算法对采集到的轨枕裂纹图像能进行更为准确的检测,裂纹边缘特征的完整性得到更好的保留。同时得到评价指标:F权重值(F-weighted)为0.831,平均绝对误差(MAE)为0.015 7,AUC (Area Under the Curve)值达到0.945 3,与其他网络模型相比,具有更好的识别性、较高的效率与鲁棒性。
A cascaded convolutional neural network-based rail crack identification algorithm was proposed to address the problems of low efficiency and poor accuracyof traditional rail crack identification and detection methods.The algorithm mainly contained the crack localization detection network and the crack saliency segmentation network,where the crack saliency segmentation network contained the crack coarse saliency segmentation module CEDNet(Crack Encoder Decoder Network) and the crack boundary refinement module CRRNet(Crack Residual Refinement Network) to realize the accurate segmentation of the rail sleeper cracks.In ballastedtrack-bed images,the crack edge features of concrete rail sleepers were easily submerged in the edge noise of ballast particles,so the rail sleeper area was extracted by gray-scale projection method before inputting into SSD network,and then the cracks on the rail sleepers were precisely located and segmented by SSD network.To further extract the complete information of the crack edge features,a significance segmentation algorithm was applied to the segmented crack regions,which was input to the CEDNet module to obtain the crack coarse significance prediction map first.Then,the prediction map was fed into the CRRNet module,where the edge information and local regions were refined to preserve the integrity of the edge features.An integrated loss function of crack detection with a mixture of BCE,SSIM,and IOU losses was also used to evaluate the deviation of cracks extracted by the crack saliency segmentation network from the real cracks.The experimental results show that using the proposed comprehensive loss function and algorithm can achieve more accurate detection of the collected rail sleeper crack images,and the integrity of the crack edge features was better preserved.The results of the evaluation indicators show that the F-weighted value(F-weighted) was 0.831,the mean absolute error(MAE) was 0.015 7,and the AUC(Area Under the Curve) value reached 0.945 3,thus indicating better recognition,higher efficiency and robustness as compared with other network models.
作者
汪晨曦
李立明
柴晓冬
郑树彬
童千倩
WANG Chenxi;LI Liming;CHAI Xiaodong;ZHENG Shubin;TONG Qianqian(School of Urban Rail Transportation,Shanghai University of Engineering Science,Shanghai 201620,China)
出处
《铁道科学与工程学报》
EI
CAS
CSCD
北大核心
2022年第6期1559-1567,共9页
Journal of Railway Science and Engineering
基金
国家自然科学基金资助项目(51975347,51907117,12004240)
国家级大学生创新项目(202110856031)
上海市级大学生创新项目(cs2010006)。
关键词
机器视觉
显著性检测
边缘提取
卷积神经网络
computer vision
saliency detection
edge extraction
convolutional neural networks
