Research On Bridge Crack Segmentation Algorithm Based On Generative Confrontation Network

As the transportation hub of the world,bridge is the basic guarantee for commercial exchanges around the world.Affected by the adverse environment and limited by the service life of bridge materials,in order to ensure the safety of transportation and people’s life and health,it is very important to carry out regular bridge disease detection.Bridge crack is one of the most common bridge diseases and it is necessary to detect bridge crack.Bridge cracks are diversified in shape,easily disturbed by noise,and complicated in background.Due to its advantages of accurate recognition and strong anti-noise ability,the deep learning method is widely used in different fields.Therefore,using deep learning method to target detection or semantic segmentation of bridge crack images has become a hot method in the field of bridge disease detection.The generative adversary network is an important branch of the deep learning method,which can generate images close to the distribution of real images by random noise.A new method of bridge crack segmentation based on generative adversary network is proposed by studying the existing model of generative adversary network.The research content is as follows:(1)In order to solve the problem of obstacles in the real crack image which contains many influence recognition models to detect and segment the crack location information accurately,a method of crack image restoration based on the generative adversary network is proposed.The method proposes a distance weighted mask to improve the model repair accuracy The VGG auxiliary discriminator is introduced in the discriminator structure to assist the original discriminator in the generative adversary network to make more accurate judgment on the input images,and reduces the training difficulty of the generative adversary network and improves the generative performance of the generative adversary network.Global loss is added to optimize the visual effect of the repaired image,and a model of generative adversary network for bridge crack image is established to reduce the possibility of texture noise problem and enhance the generative ability of diverse crack image.In the repair process,the information of the obstacle is erased and the diversified crack image is generated by using the well-trained generative adversary network,then the crack image and the crack image are respectively covered with the distance weighting mask.Next,the pixel loss,semantic loss and VGG auxiliary loss corresponding to the generative crack image are calculated,and the sum of the losses is solved.The sum of loss is sorted from smallest to largest,and the generated crack image with minimum sum loss is considered as the optimal complement map,then the area corresponding to the missing area of the repair block is spliced with the repair block.The stitched image is input into the discriminator again to calculate the global loss,and finally the overall loss is optimized to obtain the final repair result.(2)Aiming at the problem that the characteristics of small bridges crack are not obvious and susceptible to noise interference,which lead to low accuracy by using the traditional segmentation model,a small bridge cracking method based on the generated adversary network is proposed.This method introduces the segmental branch in the discriminator structure,and combines the generative adversary network and the semantic segmentation network.Simultaneously,it has the function of image reconstruction and segmentation based on super-resolution.This method adds segmental loss to make the generated super-resolution coarse crack can be learned more easily and accurately segmented by the segmental model;the traditional discriminator structure is replaced by the difference discriminator,the potential correlation between the crack image of the real small bridge and the crack image of the super-resolution coarse bridge is explored thoroughly,and the ability of super-resolution image generation of the model is improved.In addition we design a super-resolution image generative model that allows the model to generate more realistic crack images.When dealing with the crack segmentation problem of small bridges,the method first converts the low-resolution small bridge crack image into the super-resolution coarse bridge crack image,and then divides the converted super-resolution image.(3)In view of the lack of systematic evaluation system for bridge crack damage,a deep learning based evaluation method for bridge crack damage is proposed.The method combines the generative adversary network with the squeeze network for the crack damage degree evaluation field,then establishes a set of crack damage degree evaluation system,finally designs the crack image generated model for the bridge crack image,and improves the classification accuracy of the crack classification model,so that the evaluation system can evaluate the degree of crack damage more accurately.Through the research on crack image repair,small crack segmentation and crack damage evaluation,a set of crack image research system based on deep learning is established,which realizes the intelligent treatment of bridge cracks.