Research And Application Of YOLO Network In Welding Seam Detection Algorithm

Weld grinding is an indispensable process for welding casting products.The polished welds are not only more beautiful and durable,but also have better stress effects.However,the welding seam grinding workshop is often full of strong light and dust.Long-term operation in this environment will have a great impact on the heart and lung health of workers.Therefore,automatic welding seam grinding has extremely high industrial value.Welding seam inspection is the primary task of realizing the automation of welding seam grinding.The detection method based on machine vision is the current mainstream target detection method.The feature extraction of the weld image collected by the camera is carried out by using image processing algorithms or deep learning and other means,and the detection is performed based on the acquired feature information.The detection speed is fast and efficient.,Can operate independently without manual intervention.Therefore,this paper chooses the visual inspection method based on deep learning to study the weld inspection.Aiming at the problem of poor detection effect of YOLO v3 algorithm on densely distributed welds and high miss detection rate,this paper proposes two improved methods.(1)Use the DenseNet network with stronger feature extraction capabilities,higher utilization,and faster inference speed to replace the Darknet-53 network for feature extraction.(2)Add an attention mechanism to the residual module of Darknet53,and use deep separable convolution to speed up network calculations.In view of the large regression error of YOLO v3 boundary,CIoU is used to improve the distance measurement method when calculating YOLO loss.Aiming at the problem of the unbalanced ratio of positive and negative samples in YOLO v3,Focal Loss is used to improve the classification loss function of YOLO v3.In order to verify the detection effects of the two improved methods,experiments were carried out using self-built weld seam data sets,and the detection effects were compared with the original YOLO v3 algorithm.Experimental results prove that the two improved methods have improved the detection effect of the algorithm to varying degrees.Among them,the accuracy of the improved method based on the depth separable convolution and attention mechanism increased by 3.3%,and the accuracy of the improved method based on DenseNet increased by 3.9%.