Research On Steel Classification And Defect Detection Methods

In the production process,due to the difference in processing technology and raw materials,the steel surface will have different degrees of defects.The existence of defects will not only affect the aesthetics of the product,but also affect the working strength,wear resistance,and anti-oxidation and anti-corrosion capabilities of the product.Appropriate selection should be made according to the difference,and it can continue to be used when the degree of defect meets the required requirements.As a relatively common basic construction material,steel plays an important role in the stability of the entire industrial product.Under this background,this thesis chooses steel as the research object.In actual industrial production activities,the use of unqualified steel will cause serious harm,so it is necessary to carry out strict quality inspections of steel products.The existing surface defect detection methods include manual visual inspection methods and physical inspection methods.The first type of method is greatly affected by human factors,is highly subjective,and is easy to miss detection;the second type of method requires professional equipment and limited detection materials.Therefore,this thesis adopts the steel defect classification method using a neural network and the defect feature feedback fusion detection method,conducts in-depth research on steel classification and defect detection,and briefly handles the subsequent classification.The main contents of this thesis are as follows:(1)Open-close transform multi-noise denoising and super-resolution adversarial data preprocessing methodIn the given image,almost every type of image data has noise,which affects the accuracy of subsequent classification operations and the accuracy indicators of detection operations.Therefore,aiming at the situation where the small target in the complex background is easily occluded or submerged by other objects or noise,an open-close transformation algorithm is proposed to eliminate or weaken the background and noise.Given the low number of defective target samples in the dataset used,a super-resolution generative adversarial neural network(SRGAN)is proposed to realize sample data expansion based on retaining high-frequency information of defects.Combined with the original classic image data amplification methods,such as rotation and staggered cutting,it avoids the situation that the original image is unreal due to cropping or brightness changes,and realizes sample data amplification.(2)Horizontal double residual classification method for steel defectsIn the classification of surface defects,analyzed and compare the strengths and weaknesses of machine learning and deep learning classifications,and analyze the characteristics of existing residual modules.Combining the benefits of existing residual modules,a horizontal double residual module is proposed as a classification method.The core of the network,making it more suitable for classification networks with deeper layers;With the original detailed classification network framework,it is difficult to achieve relatively good classification results for a particular dataset,and the improved neural network framework combined with the horizontal double residual module proposed in this thesis is used for experimental operations.The improved neural network framework has also improved inaccuracy,and is better than the original neural network classification framework Resnet;combined with the open-close transformation method and the super-resolution generative adversarial network operation processing data samples,the final classification network accuracy reaches 99.7%,which is 5.7 percentage points higher than the original classification network,and the detection speed reaches 35 FPS.(3)Steel defect feature feedback fusion detection methodIn the defect detection task,a feature feedback fusion method combining the horizontal double residual classification method of steel defects and the addition of Spatial Pyramid Pooling(SPP module)is used to reconstruct the detection method to detect steel defect data.Improve and redefine the activation function used in the network;introduce a new loss function to solve the problem of category imbalance,and rely on the COCO data set detection indicators to evaluate and analyze the algorithm,and obtain the results of each type of defect.Average detection accuracy(AP value).The test results are analyzed and summarized,and the detected defects are graded and studied in combination with the national standard documents.The experimental results show that the introduction of variable convolution enhances the network to adapt to defects of different shapes,and the detection algorithm combined with the SPP module also improves MAP after training.The average detection accuracy of the improved detection algorithm reached 75.4%,compared with the detection before the improvement.The network improved by 2.53 percentage points,which is 9.57% higher than using the initial detection network,the single-type AP value reaches 95.42%.The detection speed reached29.72 FPS,which met the requirements of steel detection tasks in terms of detection accuracy and detection speed.