Research On Machine Vision Welding Spot Location Based On Object Detection Algorithm

Resistance spot welding is the main welding method of vehicle body,there are many welding spots in vehicle body,the quality of welding spots affects the quality and mechanical properties of vehicle body,and then affects the performance of the whole vehicle,therefore,it is necessary to detect the quality of welding spots to ensure whether the body quality meets the requirements.The way of welding spot quality detection is mainly to align the manual hand-held ultrasonic flaw detector with the center of the welding spot,and judge whether the welding spot quality is qualified through the echo signal now.With the development of machine vision,automatic welding spot quality detection can be realized,but the workshop environment is complex,the accuracy of welding spot positioning is not high,and the accurate echo signal can not be obtained to judge the welding spot quality.In order to obtain more accurate center position of welding spot,this paper used the object detection algorithm based on deep learning to locate the welding spot.The main research work is as follows:1.Firstly,it expounds the relevant theoretical knowledge of deep learning and convolutional neural network.The main contents include the structure and development of artificial neural network in deep learning;Deep learning back propagation algorithm,optimization algorithm and common loss function;Convolution theory and neural network widely used in the field of classical image feature extraction.It lays a foundation for studying the application of deep learning in the field of computer vision.2.Secondly,the object detection algorithm in the field of computer vision is studied.This paper studies three object detection algorithms with different detection paradigms: Faster RCNN,a two-stage and anchor-based algorithm;YOLOv3tiny,a one-stage and anchor-based algorithm;and Center Net,a one-stage and anchor-free algorithm,different algorithms are implemented to locate the welding spot in this paper.By training on the welding spot dataset,three different welding spot detection models are obtained,among which YOLOv3 tiny has the fastest detection speed and Centernet has the advantage in detection accuracy.3.Finally,in order to further improve the accuracy of welding spot positioning and minimize the false detection of welding spot,based on the improvement of yolov3 tiny,this paper proposes the attention detection model of welding spot,and carries out welding spot positioning by integrating the attention mechanism.According to the characteristics that the welding spot size of the welding spot data set is small and medium-sized,the feature map with higher resolution is used as the input of the model detection layer,and the attention mechanism is introduced into the feature extraction network to suppress the complex background features and reduce the impact of the background on the welding spot object detection,at the same time,EIOU is used as the coordinate regression loss function to obtain a higher quality of detection frame.The results of experiments show that the final improved model CEUP＿YOLOv3tiny compared with YOLOv3 tiny algorithm,the object detection evaluation index AP0.75 increased by 4.30 percentage points,AP0.85 increased by 16.70 percentage points,AP0.85:0.95 increased by 4.15 percentage points,while the model size decreased by73.80%.This method has great application potential for vehicle body intelligent manufacturing based on machine vision.