Research On Visual Detection Method Of Steel Plate Surface Defects Based On Deep Learning

Steel plate is an important raw material for steel production and is extensively used in aviation,manufactories and many other fields.Nevertheless,in actual production,many kinds of defects will appear on the steel plate surface,these defects tremendously influence the property of the product and grievously reduce the service life of the product.At present,most enterprises still use artificial vision detection methods and traditional machine vision detection technology to detect steel plate surface defects.These methods have low detection accuracy and poor real-time performance,which are difficult to meet production requirements.Consequently,it is pretty imperative to develop surface defect detection technology with high precision rate and fast speed.In this context,a visual detection method for steel surface defect based on the combination of image processing and deep learning is proposed,and defect image preprocessing,defect classification,defect localization detection algorithm is studied.The main research contents and epilogues contain the following four parts:(1)Steel plate surface defect image preprocessing.Five data enhancement methods were used to expand the image set of steel plate surface defect and increased the data samples for deep learning model training.A ROI detection algorithm based on the average projection gray value of the steel plate surface image was proposed.The detection rate of the algorithm was97.43%,and the average detection time was 0.004 s.The improved adaptive median filtering(IAMF)algorithm was used to remove the image noise.Combined with the experimental analysis,the IAMF algorithm was better than the adaptive median filtering(AMF)algorithm in removing the noise and keeping the edge detail information.The PSO-Gabor filter algorithm was proposed to promote the defect image.The experimental results showed that the algorithm can highlight the defect characteristics and advance the image quality.(2)The Alex Net network was reasonably improved and the steel plate surface defects were classified.Transfer learning was adopted to share Alex Net network parameters,sequeeze-andexcitation(SE)was added to the network to effectively learn important features,depthwise separable convolution was used to replace part of convolution layer,batch normalization(BN)layers were added to speed up the convergence of the network,residual network structure was adapted to avoid gradient explosion and disappearance,the Leaky Re LU function was used as the activation function to solves the problem that some neurons cannot be updated during training process.The experimental results indicated that the classification accuracy of the improved Alex Net network for steel plate surface defects was 99.44% and was increased by18.27% compared to the traditional Alex Net network.(3)The Faster RCNN network was optimized and the defects of steel plate surface were located.The IOU K-means clustering algorithm is used to obtain the number and size of anchors that are more suitable for the the steel plate surface defects;The improved Alex Net network was used for feature extraction to obtain feature maps;ROI Align was used for pooling to reduce the deviation caused by ROI Pooling.The experimental results showed that the detection accuracy of the optimized Faster RCNN network for steel plate surface defects was 0.918,and was improved by 0.115 compared to the traditional Faster RCNN network,the optimized Faster RCNN network is better able to detect complex,small-size,and slender defects in the steel plate surface compared to the traditional Faster RCNN network.(4)The experimental platform for visual detection of steel plate surface defects was designed,including the steel plate motion control part,the machine vision system part and the computer software part.The steel plate motion control part adopted servo motor and linear module to drive the steel plate to move in two directions,and controled the speed of the servo motor through PWM pulses.The hardware model and structure scheme of the machine vision system were determined.An online detection software system for steel plate surface defects had been developed,which realized functions such as image acquisition,defect detection,and defect information storage.The experimental system was tested online.The artificially manufactured surface defect images of steel plates were collected by the machine vision system.After manual annotation,these handmaked defects were sent to the deep learning model for training and the deep learning model was recored in the defect detection system,(average precision)m AP and(Frame rate per second)FPS indexs were applied to assess the detection performance of the defect detection system.The experimental results indicated that the detection m AP of the detection system was 0.93,and the FPS was 18 f/s,which realized the online detection of steel plate surface defects under complex background.