Research On Steel Surface Defect Detection Based On Deep Learning

Iron and steel has a wide range of uses in all walks of life,including but not limited to construction,aerospace,machinery,automotive.Steel surface defects can not only affect the appearance of steel surface,but also damage wear resistance,high temperature resistance,corrosion resistance and fatigue strength.A large number of steel surface defects will lead to the rejection of customers,which means a significant economic loss for the production enterprises.Therefore,the detection of steel surface defects is very important to improve the quality of steel production.This paper first introduces the research background of the steel surface defect detection and the research status of surface defect detection at home and abroad,points out the existing steel surface defect detection method,including methods based on statistical,methods based on spectral,methods based on the model and methods combined with machine learning,further introduces the research status of deep learning,while introduces the research content and structure of the paper,It is clear that this paper mainly studies the detection of steel surface defects based on the object detection network in deep learning.Then,the paper introduces the deep learning basic knowledge related to object detection,including the origin of deep learning,the basic structure and function mode of deep convolution neural network as well as the development and evaluation of object detection,the steel surface defect data sets used in the project are analyzed in the size of the defect quantity and scale,The training set and test set are divided and the training set is cropped and other data augmentation operations are carried out.According to the characteristics of the data set,the Faster RCNN network is used for detection,and each module and corresponding adjustment of the Faster RCNN are introduced.Compared with Cascade RCNN,Retinanet and YOLOV4,it has better comprehensive performance.Then,the feature extraction part of Faster RCNN network is improved.ResNet 50 feature extraction network is replaced by the introduced Res Next,Res2Net,Res Nest and RegNet feature extraction networks respectively.,whose accuracy is compared to find that RegNet feature extraction network achieves best performance.Finally,Transformer spatial attention module,channel attention SE module and mixed attention CBAM module are introduced and added to the RegNet feature extraction network respectively.The comparison proves that Transformer spatial attention module can improve the precision of steel surface defect detection better.In addition,the training strategies of transfer learning,multi-scale training and cosine learning rate are used to improve the detection precision,finally improving 3.8% of m AP.Finally,the software interface of the steel surface defect detection system is designed,and two working modes of training and testing are realized.It can train the model and predict the steel surface defect,as well as view the historical detection records.The improved Faster RCNN designed in the project can be better applied to the steel surface defect detection.