基于深度学习的混凝土裂缝检测发展综述

A Review on the Development of Deep Learning Based Concrete Crack Detection

在人工智能快速发展的背景下,计算机视觉技术进入了新的阶段,基于深度学习的混凝土裂缝检测方法可有效提高检测的准确率和精确度。语义分割具有像素级分类的特点,其在混凝土裂缝检测工作中具有较强适用性。本文对基于FCN、U-Net和ResNet架构的深度学习研究成果进行了归纳总结,分析发现通过引入图像预处理与形态学后处理相结合的方法,C-FCN像素准确率、召回率和交并比相较于FCN分别提升5.61%、16.56%、13.22%;采用Dice Los加交叉熵的混合损失函数可有效应对前景背景像素样本失衡的问题,U-Net的IoU和F1优化后较之前分别提高5.41%和5.19%;残差块短路机制可改善梯度传播、加速训练过程以及提高模型的性能,含有残差块AcNet的F1分别较UNet和RCF提高了1.17%和2.43%,OR则分别提高了1.35%和2.78%。卷积神经网络模型搭建相对灵活,根据实际情况可从参数更新方式、激活函数、损失函数等方面对其进行优化。本文对深度学习模型建立、数据集收集和模型评估三个重要部分的内容进行了总结,以期为后续相关研究提供一定帮助。

Under the background of rapid development of artificial intelligence,computer vision technology has entered a new stage,and the concrete crack detection method based on deep learning can effectively improve the accuracy and precision of detection.Semantic segmentation has the characteristics of pixel-level classification,which is analysed and found to have strong applicability in concrete crack detection work.In this paper,the research results of deep learning based on FCN,U-Net and ResNet architectures are summarized,and found that through the introduction of the combination of image preprocessing and morphological post-processing,the pixel accuracy rate of C-FCN,the recall rate,and the intersection-and-comparison FCN are improved by 5.61%,16.56%and 13.22%,respectively.The use of a hybrid loss function combining Dice Loss and cross-entropy can effectively address the issue of imbalance between foreground and background pixel samples.After optimization,the IoU and F1 of U-Net are increased by 5.41%and 5.19%respectively compared to before.The residual block short-circuiting mechanism improves the gradient propagation,accelerates the training process as well as improves the performance of the model,which contains residual blocks AcNet improves the F1 by 1.17%and 2.43%and the OR by 1.35%and 2.78%compared to U-Net and RCF,respectively.Convolutional neural network model building is relatively flexible,and it can be optimised in terms of parameter update method,activation function,loss function,etc.According to the actual situation,this paper summarises the contents of the three important parts of deep learning model building,data set collection and model evaluation,with a view to providing some help for subsequent related research.