Study On Bridge Disease Detection Based On Attention And Wavelet Transform

Bridge disease is an outstanding problem existing in our country at present.And how find the bridge disease is important.In this paper,the research status of bridge disease detection is studied in detail,and it is found that the current bridge disease detection methods are mainly artificial inspection.With the development of deep learning,object detection models are used to design automated bridge disease detection models.However,these models still have the following outstanding problems: 1.When the resolution of the image is low or the color of the image is changed under the influence of illumination,it cannot work well.2 The existing method only applies the existing object detection model to bridge disease detection,and does not study and design specific algorithms according to the task and data characteristics of bridge disease detection.It has the problem of poor detection effect and slow speed.This paper uses computer vision,deep learning,image processing and other related technologies to improve the existing target detection algorithms,and makes targeted research on the problems of "illumination influence" and "low resolution" existing in the image.The main work is as follows(1)Image illumination estimation model.In this paper,a deep neural network based on separation convolution is proposed to extract illumination from images.The model implements an end-to-end solution.It has 11 convolution layers,except the first and last ones,which are separated convolution layers.At the end of the network,"global average pooling layer" is used instead of "fully connection layer" to reduce the number of parameters of the network.The illumination estimation error on the Color Checker Dataset is about 60% lower than state-of-the-art,and the size of the model is reduced from 35.62 MB to 0.62 MB,so that the model can be well applied to mobile devices.(2)Image super-resolution reconstruction model.In this paper,a multi-scale RCAN neural network based on wavelet loss is proposed on the basis of the existing RCAN.The main improvements are as follows: 1.The multi-scale feature extraction branch is added to the residual module of the main body of the network.A larger kernel size is used in this branch.2.The wavelet loss function is designed.The difference of the high frequency coefficients between the output image and the real image after wavelet transform is calculated as the loss value,which is similar to the perceptual loss and can recover the edge details of the high resolution image better.The results are improved compared with RCAN on Set5,Set14,B100 and Urban100.(3)A new object detection model.In this paper,an object detection model based on Attention mechanism and wavelet transform is proposed.The main improvements are as follows: 1.Firstly,turn the image to gray,then use the result of wavelet transform as input.2.The Attention between feature maps is introduced into the residual module of the backbone network,which can make the network pay attention to the features that are more useful for disease detection,such as the texture feature of the image and so on.In order to verify the effectiveness of the above-mentioned models,this paper makes a comparison on the self-built bridge disease dataset with the existing target detection model Faster RCNN,RetinaNet,SSD,YOLOv3.The final result of our model in mAP is 81.92%,which is much higher than that of RetinaNet,SSD,YOLOv3 and other models.Though it is smaller than 85.62% of Faster RCNN,but the detection speed of the model is only about 50% of Faster RCNN.