| With the progress of science and technology,machine vision has been widely used in production and life,but it is rarely used in express packaging defect detection.Identification of packaging defects by machine vision inspection methods can reduce labor costs and improve inspection efficiency.Based on the deep learning model architecture,this paper proposes a packaging defect detection model based on machine vision,and further explored its application in practice.The main research contents are as follows:(1)Set up packaging defect data set.Corrugated express boxes were selected as the research object of packaging defect detection,and classified according to the shape and cause of defects.2356 defect images were captured,and Labelimg tool was used to label the defects.(2)Proposed a packaging defect detection model based on Swin-YOLOv5 s.YOLOv5s was improved to detect packaging defects with small targets and concentrated location distribution: By replacing the backbone module of YOLOv5 s with Swin Transformer,the model’s defect feature extraction and small target perception ability can be enhanced.EIo U was used to replace CIo U loss function to improve the ability and speed of model regression prediction.The algorithm was trained and tested by the self-made packaging defect data set,and the detection accuracy was 93.9%.The recall rate is 77.3%.When tested on NVIDIA Ge Force RTX 3090,the detection speed reached 35.6FPS,which has a good detection effect.(3)The sorting system was designed and the packaging defect detection model was further optimized to meet the actual application requirements.After analyzing the practical application characteristics of the packaging defect model,the packaging sorting system was designed.In order to further improve the practical application ability,the Swin-YOLOv5 s model was improved: Firstly,data augmentation is used to improve the detection effect of the model.Then,the model was pruned and quantized,and Tensor RT was used for network compression to reduce the model size and improve the detection speed.The size of the optimized Swin-YOLOv5 s model was reduced to 11.4MB and the detection speed reached 36.3FPS.The model realized the accurate detection and real-time detection of packaging defects,and has certain practical application value. |
