Research On Visual-based Knot Detection And Defect Recognition Of Wire Rod

Before leaving the factory,the wire rod of the high-speed wire plant of the steel enterprise needs to be bundled with an automatic bundling machine through wire ties.When bundling,the knots often appear unknotted and partially knotted incompletely,bringing inconvenience and safety hazards to storage and transportation.To work more accurately,efficiently,and safely,the detection and defect identification of wire rod knots have profound practical significance.At present,due to the high temperature of the strapping station and the continuous working of the wire rod strapping machine for 24 hours,the production workshop lacks an online detection link for wire rod kinks.In addition,the background of the collection of wire rod kinks images is very complicated,and the bundled wire rod has a certain angle of deflection,as well as the size and diversification of the knots.However,the traditional image detection algorithm is less robust,and it is difficult to identify the knots of the wire rod with a complex background.For wire rod knot detection and bundle quality identification,this thesis develops a deep learning-based method to detect wire rod knots and identify quality so that the detection effect can reach the ideal range.This thesis does the main work as follows:(1)According to the actual situation of the wire rod strapping machine production line,a hardware imaging system and a software detection system for wire rod knot detection and recognition are designed.And then,the bundled wire rod knots are collected several times in the actual scene.The brightness classification and adaptive Gamma transformation of the images collected under complex illumination are carried out to construct the wire rod knot data set.(2)Deeply studied the network structure of the YOLOv3 algorithm,this thesis proposed a knot detection algorithm for wire rods based on YOLOv3.Two improvements for its shortcomings were also put forward: using the lightweight convolutional network Mobile Net V3 as the feature extraction network and using CIoU?Loss to optimize the regression box loss function.Later,a comparative experiment was carried out using the real-scene collected and annotated wire rod knot data set.The experiment showed that the size of the improved network model was only 14% of the YOLOv3 model,and the average accuracy rate(Average Precision,AP)was 95.43%.The single image prediction the time only needs 0.51 s.The improved algorithm greatly improves the speed of training and detection while also increasing the average accuracy.The improved YOLOv3 algorithm has a faster detection speed for medium-sized wire rod knot images,reducing missed detection.The enhanced model can quickly detect the unknotted wire rod and extract the image of the knot area to meet the real-time requirements of the scene.(3)Use HOG+SVM,CNN classification and YOLO algorithm to conduct experimental analysis on the defects of wire rod knots.Aiming at the problem that small target "knots" are challenging to identify,an improved YOLOv5 algorithm is finally proposed to identify wire rod knot defects.Using knot images and annotations to conduct comparative experiments on the YOLOv5 algorithm and enhanced algorithm,the experiment shows that the enhanced YOLOv5 algorithm has a better recognition effect on small target "knots".The improved network regression box loss convergence is minor,and the average accuracy rate is 99.5%,which makes the knot's location more accurate,and the prediction time of a single image is only 0.22 s.The enhanced algorithm can quickly detect the number of knots,thereby identifying defective wire rod knots.