Research On Steel Plate Surface Defect Detection Algorithm Based On Deep Learning

Steel products are an indispensable part of the metal products we use in our daily lives,therefore,the quality of steel is particularly important.In the process of producing steel plates,due to various factors,the produced steel plates will have many defects,especially on the surface of the steel plates,which will affect the quality and performance of the steel plates.Therefore,it is necessary to accurately and efficiently detect these defects and solve the problems in a timely manner.Compared to traditional detection methods,there are many drawbacks,such as inaccurate detection,low efficiency,and slow detection.This article focuses on these issues and conducts research on deep learning based steel plate surface defect detection algorithms.The algorithm is optimized,which not only achieves intelligent detection of steel plate surface defects but also improves the detection accuracy and robustness of the algorithm.This article analyzes the defects on the surface of steel plates,which have the characteristics of being small and multi category.There are many types of defects,resulting in similar colors,shapes,and other similar phenomena between classes.There are also significant differences between defects in the same class.Establish a dataset of surface defects on steel plates,first enhance the dataset to obtain a sufficient number of images,and then annotate the data.Specifically analyze the feature extraction network,candidate region generation network,and classification regression network of the Faster R-CNN(Regions with CNN features)object detection algorithm.Based on this algorithm,train and learn the steel plate surface defect dataset using VGG16 and Res Net50 as the basic feature extraction networks.Analyze and compare the experimental results,This article uses Res Net50 as the basic feature extraction network for steel plate surface defect detection algorithm,and the experimental results are not ideal.Therefore,the algorithm needs to be optimized to achieve the desired effect.On the basis of Faster R-CNN algorithm,the intersection ratio of model parameters,activation function and other parameters were selected and optimized.Due to the presence of many small defects on the surface of steel plates,the problem of missed detection may arise.This thesis proposes a feature fusion and multi-scale detection network,which introduces Feature Pyramid Networks(FPN)on the basis of Res Net50 network to replace the original feature extraction network,allowing the network to obtain more information and avoiding the problem of missed detection of small defects.The original algorithm uses the Ro I Pooling method,which has the disadvantages of inaccurate positioning and low detection accuracy,mainly due to its quantitative rounding operation.In this thesis,Ro I Align is used instead of Ro I Pooling,and bilinear interpolation is used to obtain the image values of pixel points,effectively improving the detection accuracy.The experimental results show that the Faster R-CNN algorithm using feature fusion,multi-scale detection network,and Ro I Align pooling method achieves an average accuracy(m AP)of 4.84% higher than the original Faster R-CNN algorithm for detecting steel plate defects,indicating that the optimized algorithm is more accurate and efficient for detecting steel plate defects.The method proposed in this article has strong feature extraction ability and high detection accuracy for surface defects on steel plates,without the need to set a specific feature extraction algorithm for a specific defect.It is suitable for online detection of steel plate defects.