Deep Convolutional Neural Networks for Image Reconstruction and Damage Recognition in UAV Bridge Inspection

In this paper, bridge image data are collected using a UAV, and the collected images are denoised and enhanced using wavelet analysis methods and nonlinear variations. A deep convolutional neural network (DCNN) is used to construct a model for the reconstruction task of the acquired bridge images and to reduce the impact of blurring and other problems generated in the process of image compression on the bridge damage identification. An improved deep convolutional neural network with step-by-step input capability is proposed, and after feature extraction by MobileNet-v2 lightweight network through deep separable convolution operation, the spine neural network is utilized to construct a decision module so that the decision-making information is fed into the fully connected layer to obtain the bridge damage recognition results. In this paper, the mean value of the peak signal-to-noise ratio of the image obtained after denoising the image using wavelet analysis is 51.27, and the mean value of the structural similarity is 0.974, which is significantly better than other denoising methods. The image reconstruction model constructed based on DCNN improves by more than 19% compared with all other algorithms, showing high accuracy and high efficiency. The bridge damage recognition model has an accuracy of more than 90% for the damage recognition of different working condition locations of the bridge, which can effectively realize the bridge damage recognition based on the images collected by UAV.