The Key Technologies Of Automated Visual Detection Based On Deep Learning

Material sorting and defect detection are the two most common operation scenarios in the industrial production field.Material sorting refers to selecting materials placed in disorder according to production requirements;defect detection refers to inspecting the appearance and performance of products or parts.Screen out defective products and improve product yield.These two tasks have high repetitiveness,a single implementation background,and low manual operation efficiency.As labor costs increase year by year,the two have become important application scenarios in the field of industrial automation production.In the industrial automation production,the target detection technology in the research direction of computer vision is the most widely used.Most target detection algorithms in traditional automated production are machine learning methods.As the target detection algorithms based on deep learning become more mature,their performance in practical applications is much better than traditional machine learning algorithms.This article focuses on the two automated production requirements put forward by related companies.These two requirements are:material sorting task,which aims to identify the scattered electronic components in the material tray,and transmit their type and location information to the robotic arm,which is picked up in order and accurately placed in the production tray.;Defect detection task,the task is to pass the panel defect status to the robotic arm,the robotic arm removes the defective panel,and the flawless panel enters the next round of production.Based on the above requirements,this paper separately evaluated the performance of four deep learning-based target detection algorithms on task data sets,SSD,YOLOv5s,Faster-RCNN and Mask-RCNN,and selected the best performing YOLOv5s algorithm as the basic algorithm.At the same time,combined The two task objectives of material sorting and defect detection are different,and the algorithm is improved respectively.The final experimental results show that the algorithm can meet the needs of the enterprise’s automated production.The automated production system includes an image processing part and a power transmission device part.This article mainly studies the target detection algorithm of the image processing part,and does not involve the circulation and operation of the entire automated production system.The research content of this article includes three aspects:(1)Making data sets and evaluating the performance of four algorithms.All target detection algorithms need data sets for training.This paper is an engineering application research based on specific tasks,and there is no public data set that meets the requirements.Therefore,the task data set needs to be collected by the author himself.The enterprise could not provide the conditions for data collection in the production environment,but only provided the corresponding materials.This paper completed the data collection and labeling work in the laboratory environment according to the task characteristics and requirements.Secondly,quantitative evaluation criteria of algorithm performance were designed for the two tasks respectively.Based on the task data set,three target detection algorithms were evaluated for each task.According to the evaluation results,the YOLOV5S target detection algorithm with the best performance was selected for both tasks as the basic algorithm for completing the task.(2)Based on the defect detection data set,the YOLOV5S algorithm was improved to improve the defect panel detection ability.Firstly,in the defect detection task,this paper proposes a secondary transfer learning method to solve the problems of small data volume,difficult data collection and restriction of algorithm training effect of data set.This method can alleviate the restriction of data set quantity.Secondly,this paper improves the anchor frame generation strategy in the YOLOV5S model to make the prior frame size closer to the actual size of the target in order to solve the problem of high false detection rate of good panel.Experimental results show that the improved algorithm can meet the production needs of enterprises and improve production efficiency.(3)based on material sorting data collection,this paper deeply analyzes the detection performance YOLOv5s algorithm,based on the algorithm,several kind of detection of defects,strengthened the data set is designed,targeted to expand the data set,in addition,according to the algorithm of adjacent weak target detection ability,often leaving out the problem of small targets,redesigned the algorithm outputs the maximum inhibition method,The ability of the algorithm to detect overlapping targets is enhanced.Experimental results show that the performance of the improved algorithm is better than that of the basic algorithm,and the phenomenon of missed and false detection is obviously reduced.