Research And Implementation Of Steel Surface Defect Detection Method Based On Improved RCNN

With the improvement of global manufacturing level,manufacturers have higher and higher quality requirements for steel products.Surface defects in steel products are one of the important factors affecting product quality.In this thesis,the surface defect of steel strip in a steel enterprise is taken as the research object,and the method of detecting surface defects of iron surface based on deep learning is designed and realized.It has achieved good results and has a good guiding effect on the surface quality assessment of strip steel enterprises.The main research contents are as follows:(1)The research status of strip surface defect detection is reviewed.The advantages and disadvantages of current mainstream defect detection methods are analyzed.The causes and characteristics of strip surface defects are summarized.The working principle and several kinds of convolutional neural networks are introduced.Typical up-to-date deep learning network and target detection network.(2)A method for detecting surface defects of strip steel based on improved Faster R-CNN is proposed.The strip surface defect data set is constructed,and three improvements are made to the Faster R-CNN network structure for the characteristics of the surface defects of 8 typical strip products.In the MxNet deep learning environment,the Faster R-CNN network is built.The ResNet-101 four networks is used to extract the network to replace the original VGG16 network,and the deformable convolution(DCN)is added to increase the mAP value by 2%.Further,on the basis of using DCN,the Dense structure is added to the feature extraction network to extract more semantic information,and the final result is increased by 3.2 percentage points.Finally,based on the above two methods,the training process is performed.Online Difficult Sample Mining(OHEM),the final result achieved a 4.3%increase in the best test results.The experimental results show that the proposed algorithm has a good detection effect on steel surface defect detection.(3)A method for detecting surface defects of strip steels for incompletely labeled samples is proposed.Firstly,the method of detecting surface defects of strip steel based on Mask RCNN is realized.This method uses ResNet as the feature extraction network,and at the same time realizes the detection and segmentation of strip steel product defects.Secondly,in view of the difficult problem of marking samples,a method for detecting surface defects of strips with incompletely labeled samples is proposed.This method uses an optimized iterative method without changing the network structure.The experimental results verify the feasibility of the method,and the final experimental results show that the detection effect under incomplete labeling can be close to the detection effect under complete labeling.(4)Based on the specific needs of users,a deep surface learning based strip surface defect detection system is designed and implemented.The system can realize the defect detection of the user online and synchronously output the detection result,and simultaneously perform statistical analysis on the detection result information and visually display in the form of a histogram and a pie chart.