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

With the development of manufacturing system,the application field of steel plate material is widely used,but the surface defects of steel plate seriously affect the quality of steel plate.The manufacturing industry often uses image inspection technology to identify defects such as cracks or pitting on the steel surface.After introducing deep learning into defect detection,there are still problems that the detection speed and detection accuracy do not meet the actual demand.Therefore,this dissertation uses deep learning algorithms to investigate the speed and accuracy of steel plate surface defect detection respectively.Manufacturing companies focus on production efficiency,so the speed of steel plate surface defect detection needs to meet the requirements of real-time detection rate.In this dissertation,we study a lightweight-based YOLOv4 steel plate surface defect detection method.Firstly,the feature extraction network is optimized to be lightweight,and the original CSPDark Net53 network is replaced by the Moblilenetv3 network to extract the surface defects of steel plates;Secondly,the spatial and channel attention modules are introduced in the 3×3 convolution of feature fusion to compensate for the accuracy loss caused by the lightweight network;Finally,the use of deep separable convolution in the spatial pyramid mechanism can reduce the number of parameters of the network and improve the generalization ability of the network.Experiments on the NEU-DET public dataset significantly improve the detection speed of steel surface defects compared with the existing classical model.Accuracy is a critical measurement when detecting surface defects on steel plates.Due to the different shapes of steel plate surface defects and the small area of pits are not easy to detect,which seriously affects the accuracy of steel plate surface defect detection.In this dissertation,we study a high precision Faster RCNN steel plate surface defect detection method.For the characteristics of steel plate surface defects,K-means++ algorithm is introduced to generate more suitable anchor box for steel plate surface defects;the feature extraction network adopts dense connected feature pyramid structure of residual network to improve the feature extraction ability of the model for defects;finally,DIo U is used to optimize the boundary regression prediction.Analyze the effect of positive and negative sample imbalance of steel plate surface defects,and introduce online difficult sample mining technique for network training.Designing experiments,comparing with the classical algorithm,have better detection effect on the surface defects of steel plates with small targets and complex background.The lightweight YOLOv4 steel plate surface defect detection method studied in this dissertation can reach 82.8 frames per second with an average accuracy reduction of 1.3%,and is suitable for online detection in combination with embedded and other hardware platforms.The Faster R-CNN based steel plate surface defect detection method can achieve an average accuracy of 93.7% with a speed reduction of 3.5 frames per second,which is suitable for the detection of complex background and small target defects.