Research On Keypoint-based Algorithms And Reduction Of Their Computational Complexity For Steel Defect Detection

During the smelting process,steel products may have surface defects due to improper temperature control and impure steel.Steel with surface defects may cause a safety hazard to the subsequent process and affect the quality of the steel.Therefore,it is necessary to detect steel surface defects in the production line.Most of the traditional surface defect detection methods use manual detection,which has problems such as high labor intensity and reliance on the experience of the inspector.The defect detection method based on computer vision can complete the non-destructive detection of surface defects with high accuracy and high detection efficiency.Therefore,this paper chooses the method based on computer vision to realize defect detection.Object detection algorithm based on computer vision are divided into detection algorithm based on manual feature extraction and detection algorithm based on deep learning.Since the algorithm based on manual feature extraction has problems such as insufficient ability to extract image features and low model generalization ability,this paper chooses the object detection algorithm based on deep learning.The object detection algorithm based on key points is one of the important branches of the object detection algorithm based on deep learning.Compared with the anchor-based object detection algorithm,the object detection algorithm based on keypoint does not need to manually set the size of anchors,and the detection boxes can fit the steel defect objects with various shapes.Therefore,this paper uses the object detection algorithm based on keypoint to realize defect detection,and solves the problems of discrepancy in the prediction of keypoints and high complexity of model calculation.The main research contents of this paper are as follows:In order to solve the problem of discrepancy in the prediction of keypoint in the algorithm,this paper first makes an experimental analysis of the information discrepancy between corner points,and designs a penalty function for the confidence of the detection boxes according to the embedding vector distance between corner points.Secondly,the predicted center point position and classification ability are analyzed,reorder and weight center points.Comparative experiments show that the designed confidence penalty function can reduce the number of false detection boxes.Reordering the center points and adjusting the weight,the detection mAP of CenterNet on the steel defect dataset is increased by 1.9%.In order to solve the problem of high computational complexity of the model,we design a group convolution of information fusion between adjacent channels for high-capacity model.We also design a group convolution of breadth-first information fusion between channels for the low-capacity model,and implements adaptive group convolution.Comparative experiments show that the adaptive group convolution can effectively reduce the parameters and calculation of the model.Applying adaptive group convolution to CenterNet,the parameters of the model is reduced to 12.5%,and the FLOPs(floating point operations)are reduced to 24%.