基于改进DeepLabV3+网络的坝面裂缝检测方法

Dam surface crack detection method based on improved DeepLabV3+network

基于图像分析的水电站坝面缺陷判别是一种高效、精准的方法,然而大坝裂缝图像存在背景复杂以及裂缝和背景像素比例不均衡等问题,导致传统算法的检测效果差。该文提出一种基于改进的DeepLabV3+网络模型的坝面裂缝检测方法。该方法利用三线注意力模块(three line attention module,TLAM)提高模型对裂缝像素的提取能力;采用空洞空间金字塔池化(atrous spatial pyramid pooling,ASPP)模块进行级联优化,实现更密集的像素采样,获取更丰富的裂缝特征;将MobileNetV2作为特征提取主干网络,可实现网络轻量化并减少模型参数;将焦点损失(focal loss,FL)和Dice损失(Dice loss,DL)作为模型的损失函数,可克服数据不平衡的困难。对西南某水电站的坝面裂缝数据集进行模型有效性和对比实验,结果表明:该方法中用来评价模型精度的F_(1,score)、平均交并比(mean intersection over union,MIoU)和平均像素精度(mean pixel accuracy,MPA)的值分别达到73.98%、66.73%和73.81%,比改进前的DeepLabV3+网络检测效果分别提高了3.33%、2.89%和1.12%,该方法可为坝面维护以及未来风险评估提供技术支撑。

[Objective] Image analysis is an efficient and accurate method for identifying hydropower dam surface defects.However,due to the complex background of dam crack images and the uneven proportion of cracks and background pixels,the detection effect of traditional algorithms is poor.Moreover,traditional artificial crack inspection is not only inefficient but also costly in the present day.Efficient and accurate dam surface crack detection techniques are crucial for dam maintenance and operation.In order to achieve accurate and efficient dam surface crack detection,a dam surface crack detection method based on an improved DeepLabV3+model is proposed.[Methods]Model training is carried out for the self-made dam surface crack image dataset of a hydropower station in Southwest China,and the model is evaluated by F _(1.score),Z_(Miou),Z_(MPA),parameter quantity and other indicators.The following improvements are made to the traditional DeepLabV3+network model:(1)A three line attention module is added to improve the model's ability to extract crack pixels and reduce proportion imbalance between background pixels and crack pixels.(2)The original pyramid pooling module is cascaded for model optimization so that the model can achieve more intensive pixel sampling,and subsequently obtain more abundant crack features.(3)In order to solve the problem of the significant/too large number of traditional DeepLabV3+network parameters,MobileNetV2 network is used as the backbone of the model to extract the network,to reduce the network to a lightweight module,and to reduce model parameters.(4)Focal loss and Dice loss are used as the loss functions of the model to overcome the data imbalance and to improve the accuracy of network classification.[Results]The improved DeepLabV3+network model in this paper could better realize the extraction of crack pixels,reduce the problems caused by the im balance of pixel proportion,and better ensure the efficient and accurate detection of dam surface cracks.The experiment on the self-made dam surface crack image dataset of a hydropower station in Southwest China showed the following:(1)Compared with the original model,the improved DeepLabV3+model increased F _(1.seore )by 3.33%,Z_(Miou) by 2.89%,Z_(MPA) by 1.12%,and the parameters were reduced to 3 o14714.This finding showed that the improved model proposed in this paper had stronger performance than the original model,better ability to extract crack pixels,and could better complete the task of crack identification.(2)Compared with other attention mechanisms,the three line attention module proposed in this paper had certain advantages,which could increase the attention of the model to the crack pixels and enable the model to extract the crack features needed.[Conclusions]Through an analysis of the experimental results,the model improved in this paper has stronger segmentation ability,less missing data and false detection,and can effectively complete the dam surface crack segmentation task.The improved method increases the efficiency and accuracy of dam surface crack detection and reduces the model parameters.It can provide powerful data support for crack detection and the safe operation of hydropower projects.