Research On Machine Vision Based Surface Defect Detection Method For Steel Rails

As a strategic,pioneering and important national transport infrastructure,the railway is the engine of national regional development as well as the backbone of the comprehensive transport system.As the speed,density and load capacity of railway trains increase,the friction between the rail surface and the train wheels increases,thus increasing the probability of rail surface defects.Rail defects can affect the comfort of passengers and can cause operational safety risks.Therefore,it is important that rail surface defects are accurately detected and maintained in a timely manner.The current method used for rail surface defects detection is difficult to adapt to the rail complex service environment,this thesis for the rail surface image in the noise pollution conditions is difficult to effectively reduce the noise problem,as well as stains lead to false detection and defects of different shapes lead to difficult to accurately detect the problem,and the development of rail surface defects detection system,the main content is as follows.(1)In order to solve the problem of rail surface defect image noise reduction relying on artificially set parameters and blurred defect edges,a rail surface defect image noise reduction method based on attention-guided poly-scale noise reduction convolutional neural network is proposed.Firstly,poly-scale convolution is used to extract and learn the features of the noisy image to overcome the problem of blurred edges caused by the lack of fine extraction of single-scale convolutional features;secondly,jumping connections are used to fuse deep and shallow features of the network to strengthen the influence of shallow features;then the attention mechanism is used to adjust the weights of the features at different locations in space to filter out the features that can characterize the noise and obtain the noise information;finally,the reconstruction Finally,the reconstruction module is used to remove the noise information from the noisy images to achieve end-to-end image noise reduction.The experimental results demonstrate that the proposed method not only has a better noise reduction effect but also retains the defect edge information more effectively from both qualitative and quantitative perspectives.(2)Considering the problems of false detection due to stains and difficulty in accurate detection due to different shapes of defects on the rail surface,an improved YOLOv7-based rail surface defect detection method is proposed.Firstly,a rail image containing a stain is added as a negative sample when constructing the dataset,so that the network can learn to distinguish between defects and stains by using the label difference,and overcome the problem of false detection of stains as defects during defect detection.At the same time,a channel attention mechanism is embedded in the network,and its feature of weighting the features of different channels is used to make the network take into account the importance of different channels when extracting feature information and enhance the model’s ability to extract defect features.The experimental results demonstrate the effectiveness and feasibility of the proposed method.(3)Combining the characteristics of rail surface images and machine vision defect detection process,analyzing the functional requirements of rail surface defect detection system,and carrying out the design and functional implementation of the overall scheme of rail surface defect detection system based on machine vision.The software can realize the functions of data acquisition,data management,data annotation,data set division,noise reduction model training and application,detection model training and application,detection result statistics and manual verification,which provides a reference for the practical application of noise reduction methods and detection methods.The thesis concludes with a summary of the entire work,describing the remaining shortcomings of existing methods and providing an outlook for subsequent related research.