摘要
针对移动端平台计算资源紧张,难以应用计算复杂度高的语义分割网络进行隧道衬砌裂缝实时检测的问题,提出一种基于改进金字塔场景解析网络(PSPNet)的实时分割网络模型Mobile-PSPNet,以减少模型对计算资源的需求。以改进的MobileNetV2轻量网络替换Resnet50作为主干网络以大幅减少模型复杂度,同时在深层主干网络引入适用于移动端设备的h-swish激活函数补偿因替换主干网络所损失的精度,在浅层主干网络利用卷积注意力机制(Convolutional Block Attention Mechanism)从通道和空间2个维度提升网络对裂缝特征的关注度以增强网络抗干扰性,最后采用组合损失函数从全局和局部进行损失计算以处理裂缝图像的样本不均衡问题。实验结果表明:组合损失函数能更为精准地对裂缝图像进行分割;Mobile-PSPNet在自制数据集上的交并比为73.74%,在GPU上预测单张473×473的图像耗时为26 ms。Mobile-PSPNet具有与主流模型相当的精度和更快的分割速度,更适合部署于移动端平台以进行隧道衬砌裂缝的实时检测。
In order to solve the problem of applying semantic segmentation networks with high computational complexity for real-time tunnel lining crack detection on mobile platforms with limited computational resources,a real-time segmentation model Mobile-PSPNet based on improved Pyramid Scene Parsing Network was propose to reduce the computing resources of the model. The Resnet50 was replaced with an improved MobileNetV2lightweight network as the backbone network to reduce model complexity significantly. Besides, the h-swish activation function for mobile devices was introduced in the deep backbone network to compensate for the loss of accuracy due to the replacement of the backbone network. At the same time, the Convolutional Block attention mechanism was introduced in shallow backbone network to enhance the network ’s attention to crack features from both channel and space dimensions, so as to improve the network’s resistance to interference. Finally, a combined loss function was introduced to calculation the loss globally and locally to deal with sample imbalance problem of crack images. The results show that the crack images can be segmented more accurately by adopting the combined loss function. Mobile-PSPNet takes 26 ms to predict a single 473×473 image on GPU, and the Intersection over Union(IoU) is 73.74% on the homemade dataset. With comparable accuracy and faster segmentation speed as mainstream models, Mobile-PSPNet is more suitable for real-time detection of tunnel lining cracks on mobile platforms.
作者
宋益
赵宁雨
颜畅
谭海鸿
邓杰
SONG Yi;ZHAO Ningyu;YAN Chang;TAN Haihong;DENG Jie(School of Civil Engineering,Chongqing Jiaotong University,Chongqing 400074,China;State Key Laboratory of Mountain Bridge and Tunnel Engineering,Chongqing Jiaotong University,Chongqing 400074,China)
出处
《铁道科学与工程学报》
EI
CAS
CSCD
北大核心
2022年第12期3746-3757,共12页
Journal of Railway Science and Engineering
基金
国家自然科学基金资助项目(52078090)
重庆市自然科学基金资助项目(cstc2021jcyj-msxmX1011)。