无砟道床表观伤损智能识别算法研究

Research on Intelligent Recognition Algorithm on Surface Damage of Ballastless Roadbed

为提高无砟道床表观伤损检测系统的检测精度和准确率,提出了一种多尺度多任务的伤损智能识别算法。采用特征图L1范数准则对ResNet网络的冗余卷积核进行压缩剪枝,以降低模型计算量和存储空间。通过采集的图像构建样本库,利用像素级语义分割算法,以优化后的ResNet网络为编码网络,以PPM网络为解码网络,搭建编码-解码深度学习架构模型,并通过测试集试验和现场试验对模型进行验证。结果表明,该模型对2000张测试图像的识别准确率为95.6%,无砟道床表观伤损现场检出率为96.4%,检测效果良好。该模型可以实现对无砟道床表观伤损的自动化检测、伤损趋势分析和状态评定。

In order to improve the detection precision and accuracy of the surface damage detection system of ballastless roadbed,a multi-scale and multi task damage intelligent recognition algorithm was proposed.The L1-norm criterion of characteristic graph was used to compress and prune the redundant convolution kernels of ResNet network,so as to reduce the amount of calculation and storage space of the model.The sample database was constructed through the collected images,and the pixel level semantic segmentation algorithm was used to build the coding decoding deep learning architecture model with the optimized ResNet network as the coding network and PPM network as the decoding network.The model was verified by test set and field test.The results show that the recognition accuracy of 2000 test images is 95.6%,and the on-site detection rate of apparent damage of ballastless roadbed is 96.4%.The detection effect is good.It can realize the automatic detection,damage trend analysis and state evaluation of apparent damage of ballastless roadbed.