基于改进深度学习网络对裂缝检测研究

Research on crack detection based on improved deep learning networks

提出了一种运用于混凝土裂缝检测的模型yolov5-slC,该方法建立在yolov5模型的基础上,改进yolov5s网络结构,将普通卷积替换为轻量卷积GSConv,在此基础上继续引用slimneck结构,降低网络复杂度的同时并未丧失精度,缩短检测时间,减少计算量;引入轻量化上采样算子CARAFE,减少了计算开销。与原yolov5s相比,该方法的mAP_(50)提高了3.5百分点,mAP95提高了9.4百分点。在试验数据集和补充数据集进行可视化对比,yolov5-slC均表现良好,能高效精确地为混凝土裂缝检测提供新方法。

A model yolov5-slC for concrete crack detection is proposed.This method is based on the yolov5 model.The yolov5s network structure is improved,and the ordinary convolution is replaced by the lightweight convolution GSConv.On this basis,the slimneck structure is continued to be used to reduce the complexity of the network without losing accuracy,shorten the detection time,and reduce the amount of calculation.The lightweight upsampling operator CARAFE is introduced to reduce the computational overhead.Compared with the original yolov5s,the mAP_(50)of this method is increased by 3.5 percentage points,and the mAP95 is increased by 9.4 percentage points.The visual comparison between the experimental data set and the supplementary data set shows that yolov5-slC performs well,so it can provide a new method for concrete crack detection efficiently and accurately.