摘要： 坝面缺陷检测是水利枢纽安全巡检的关键环节，但复杂环境下坝面图像存在干扰噪声大和像素不均衡等问题，造成坝面裂缝难以精细分割。提出一种利用可分离残差卷积和语义补偿的U-Net裂缝分割方法。在U-Net网络的编码端构建更大尺寸的可分离残差卷积模块替换常规卷积模块，从而扩大特征层感受野并避免丢失裂缝边界信息，同时在解码端增加语义特征补偿模块改善多尺度特征融合效果，将焦点损失函数和中心损失函数作为目标函数，加大裂缝前景与困难样本的损失权重以提高分类准确度。在自制西南某水电站坝面裂缝数据集上的实验结果表明，该方法的F1值和交并比分别达到69.89%与53.72%，分割效果较SegNet、FCN-8S等传统方法更优，对细小裂缝区域的识别能力更强。

关键词: 坝面裂缝, 可分离残差卷积, 语义补偿, 焦点损失, 中心损失

Abstract: Detection of defects on dam surface is a key part in safety inspection for water conservancy projects.However,dam surface images under complex environment are faced with large interference noise and unbalanced pixels,making it difficult to segment cracks on dam surface accurately.This paper proposes a U-Net-based method for crack segmentation using separable residual convolution and semantic compensation.At the coding end of U-Net,a larger separable residual convolution module is constructed to replace the conventional convolution module,so as to expand the receptive field of feature layer and avoid losing the crack boundary information.At the same time,at the decoding end,a semantic feature compensation module is added to improve the effect of multi-scale feature fusion.Take the focal loss function and center loss function as the objective function,the loss weights of crack foreground and difficult samples are increased to improve the accuracy of classification.Experimental results on a self-made dam surface crack dataset of a hydropower station in Southwest China show that the F1 value and Intersection over Union(IoU) of the proposed method reach 69.89% and 53.72% respectively.The segmentation effect of this method is better than that of the traditional methods such as SegNet and FCN-8S,and it has better performance in small crack recognition.

Key words: crack on dam surface, separable residual convolution, semantic compensation, focal loss, center loss

中图分类号: 

• TP391.41