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

The development of industry is inseparable from the support of steel materials,and the development of modern industry has increasingly higher requirements for the quality of steel plates.In the process of steel plate production,there will be many types of defects,such as roll printing,scratches and scarring,which will directly affect the quality and performance of steel plate,so it is necessary to effectively detect the defects of steel plate.Aiming at the problems of low automation,slow detection speed and low precision of the current traditional steel plate surface defect detection methods,this paper studies a steel plate surface defect detection method based on deep learning to realize the intelligent detection of steel plate defects.The surface defects of the steel plate have the characteristics of multiple categories,multiple scales and small targets,and the same defect has different shapes,and different defects have similarities.This paper analyzes five common steel plate surface defect samples,and establishes a sample set of steel plate surface defects based on deep learning by screening the collected steel plate defect images,and performs sample set labeling and division.The advantages and disadvantages of several deep learning frameworks are analyzed,and the Py Torch framework is selected in the experiment to train and learn the sample set of surface defects of the steel plate.An experimental system for steel plate surface defect detection is designed,and some relevant indicators for evaluating the comprehensive performance of the deep learning detection model are analyzed and explained.Through the analysis and comparison of the R-CNN and YOLO two series of defect detection models,it can be seen that the model based on the R-CNN series is a second-order detection model.Faster R-CNN,which performs better,implements two detections by introducing the RPN network,and the detection speed and precision are improved.The model based on the YOLO series is a first-order detection model,and the better-performing YOLOv3 uses residual structure and feature fusion to complete multi-scale detection at one time.The sample set of this article is tested separately on two network models.Through the analysis of the experimental results,it can be seen that the comprehensive detection performance of Faster R-CNN is better.Therefore,this article uses Faster R-CNN as the basic network model for steel plate defect detection.By analyzing the structure of the Faster R-CNN model,it is proposed to introduce the residual structure and feature fusion of the YOLOv3 model into the Faster R-CNN model,and optimize the Faster R-CNN model from the feature extraction algorithm and the RoI pooling method.The original network uses the 13-layer convolution of VGG16 to extract features.The number of network layers is small,the extracted feature map information is not perfect,and the detection precision is low.Therefore,the residual structure module is introduced to increase the number of network layers,and the down-sampling features are After the image is up-sampled,the horizontal connection and convolution are performed to fuse the features of different levels.At the same time,the hole convolution is added to optimize the residual module,expand the receptive field,and improve the resolution of the feature map,which is conducive to the realization of multi-scale defect detection with multi-layer feature maps.Due to the calculation deviation caused by RoI Pooling quantization rounding,it directly leads to the deviation of the pixels in the RoI area and the deviation of the spatial corresponding position,which affects the detection precision.In the proposed RoI Align pooling method,the characteristic value of the sampling point is obtained by bilinear interpolation of the surrounding integer characteristic points,which will not cause quantization deviation,and effectively improve the precision of steel plate defect detection.The analysis of the results of grouping and combination experiments shows that the average detection precision of the Faster R-CNN network model optimized by the feature extraction algorithm and the pooling algorithm has been improved,and after the two algorithms are optimized and combined,the average detection precision of steel defects reached 94.49%,which can meet the actual needs of steel plate production.Moreover,the steel plate defect detection method based on deep learning can effectively extract defect features without setting a specific feature extraction algorithm for each defect,which has the advantages of fast detection speed and high detection precision,and is suitable for online detection of steel plate defects.