Research On Crack Detection Technology Based On Convolutional Neural Network

Concrete plays a very important role in the construction of highways,bridges and other transportation infrastructure.However,with the use of these facilities over time,as well as the impact of rain and snow disasters,material aging and many other factors,these concrete structure facilities will inevitably suffer different degrees of damage,and among these damages,cracks are the most common For the most serious diseases,timely and effective detection of these cracks is of great significance for later maintenance.The traditional concrete crack image detection is mainly manual detection.This process not only has low detection efficiency but also requires a lot of manpower and material resources.With the continuous in-depth research on crack detection technology,nondestructive testing based on image processing has attracted wide attention because of its advantages of simple operation and high detection efficiency.However,the concrete crack image detection technology based on traditional digital image processing is mostly designed for specific scenes.Once the scene changes,the detection results are often unsatisfactory.Aiming at the problems of poor image cracking effect and weak generalization ability of traditional image processing technology,this paper deeply studies the concrete crack image segmentation algorithm based on convolutional neural network,and designs end-to-end segmentation based on ResNet101 backbone network model.It incorporates more low-level features,making the fracture segmentation results more refined.In order to be more in line with the actual application scenario,this paper also constructed a real scene crack image data set,trained a deep learning network model,and verified the model.Compared with other methods,the algorithm in this paper has higher detection accuracy and generalization ability.In practical engineering applications,people usually use metric units such as millimeters,centimeters,etc.to quantify cracks,and pixels are used in the image unit.Therefore,if you want to calculate the actual size of the crack,you must know the actual size of the unit pixel.Aiming at this problem,this paper proposes a non-contact crack measurement method based on parallel laser.By designing a parallel laser,the pixel positions of the two laser spots in the image are identified,so that the actual size represented by the unit pixel can be easily obtained.Finally,the image processing technology is used to analyze the crack segmentation results,and the actual length and width of the crack are further obtained.The pixel calibration method proposed in this paper does not require other contact-type calibration objects and the parallel laser device is compact,flexible and easy to use,and can be easily used with smart phones.Finally,this paper integrates the proposed concrete crack segmentation algorithm and parallel laser calibration method,and establishes a set of concrete crack image detection system based on Android mobile phone and server.The mobile phone combines the parallel laser to capture the image and upload it to the server for request processing.After receiving the request,the server calls the trained model to detect and segment the crack image,and the test returns the result to the client for reference by professional engineers.