结合深度学习与注意力机制的墙体安全检测模型

A wall safety detection model combining deep learning and attention mechanism

为解决传统建筑墙体检测采用人工目视方式效率低、成本高、危险性大的问题,提出利用无人机拍摄建筑外墙缺陷图像,采用卷积神经网络(Convolutional Neural Network,CNN)结合注意力机制实现对威胁建筑外墙安全缺陷的识别分类。从获取建筑外墙缺陷图像数据开始,制作缺陷图像数据集,以威胁墙体安全的缺陷为学习样本,构造浅层卷积神经网络,融入BAM(Bottleneck Attention Module)注意力机制,从卷积神经网络提取的浅层特征中提炼缺陷特征进行学习,实现建筑外墙的安全检测。经试验,多类安全问题检测正确率达到96.18%,所提出的模型相较传统的CNN、VGG 16、ResNet 18算法,检测正确率分别提高了3.36个百分点、3.92个百分点、14.6个百分点。研究表明,卷积神经网络结合注意力机制的方法可以避免局部缺陷丢失,提高检测正确率。

To solve the problems of low efficiency,high cost,and high risk in the traditional manual inspection of building walls,this paper proposes a method using UAV(Unmanned Aerial Vehicle)to capture the defection image of building exterior wall,and using CNN(Convolutional Neural Network)combined with an attention mechanism to obtain the recognition and classification of defects threatening the safety of building exterior wall.In this paper,the UAV was used to obtain the defect image data of the building exterior wall,and the defect image data set was made by cutting and adding labels.Then,a part of the defect image data threatening the safety of the wall was used as the training data set to train the model.Furthermore,the shallow convolution neural network structure was built and integrated into the BAM(Bottleneck Attention Module)model.The BAM was regarded as the last layer of the convolution layer.The defect features extracted from the shallow features extracted by the convolution neural network were used for learning and network parameter optimization.Higher weights were assigned to important defect features for obtaining the safety detection of the building exterior wall.The test set consisting of untrained defect image data was used to verify the experiment.The experimental results are as follows:the second class wall safety detection accuracy rate reaches 97.68%;the multi-class wall safety problem detection accuracy rate reaches 96.18%;The detection accuracy of this paper is 3.36%higher than that of the traditional CNN model,which shows that the convolutional neural network integrated with BAM can extract more abundant defect features and improve the accuracy.The accuracy of this model is 3.92%and 14.6%higher than the algorithms of excellent vgg16 and resnet18,respectively,and the fluctuation of accuracy is the smallest during the training process.The results show that this model can not only reduce the network structure but also improve the generalization ability of wall safety detection.The experimental results show that this method can avoid the loss of local defects and improve detection accuracy.