摘要
针对现有桥梁裂缝语义分割模型泛化能力弱、存储资源需求大等问题,提出了一种改进的Deep Labv3+模型,该模型将特征提取网络替换为轻量化网络Mobile Netv2,并结合了Swish激活函数和迁移学习策略。为验证改进模型的有效性,利用复杂背景干扰下不同类型的桥梁裂缝图片构建桥梁裂缝数据集,对改进的Deep Labv3+模型、Deep Labv3+模型、Segnet模型和Unet模型进行裂缝识别训练,从分割精度、平均交互比及模型大小等方面对四个模型的识别效果进行对比分析。分析结果表明,改进的Deep Labv3+模型分割精度达到93.41%,平均交互比达到78.51%,F1分数达到83.60%;改进模型大小仅为6.64MB,与Segnet模型大小处于同一数量级水平,明显小于Deep Labv3+和Unet模型。
In order to solve the problems of weak generalization ability and large storage resource demand of existing bridge crack semantic segmentation models,an improved DeepLabv3+model was proposed,which replaced the feature extraction network with the lightweight network MobileNetv2,and combined with the Swish activation function and transfer learning strategy.In order to verify the effectiveness of the improved model,a bridge crack dataset was constructed by using different types of bridge crack images under complex background interference,and the cracks of the improved DeepLabv3+model,DeepLabv3+model,Segnet model and Unet model were trained for crack recognition,and the recognition effects of the four models were compared and analyzed from the aspects of segmentation accuracy,average interaction ratio and model size.The analysis results show that the segmentation accuracy of the improved DeepLabv3+model reaches 93.41%,the average interaction ratio reaches 78.51%,and the F1 score reaches 83.60%.The size of the improved model is only 6.64 MB,which is at the same order of magnitude as the size of the Segnet model and significantly smaller than that of the DeepLabv3+and Unet models.
作者
陈舟
温嘉豪
卢汉文
杜和坪
CHEN Zhou;WEN Jiahao;LU Hanwen;DU Heping(School of Civil Engineering and Transportation,Foshan University,Foshan 528225,China)
基金
科技部高端外国专家引进计划项目资助(G2022030014L)
