Defect Detection Of Plastic Labels Based On Deep Learning

Automated defect detection is crucial for the quality control of products in advanced manufacturing and has been widely used to detect anomalies and defects in various products during production.As a unique identification and quality indicator,plastic label shows the information of the product,the design concept of the manufacturer,and also represents the quality of the product.Therefore,it is very necessary to detect and identify the surface defects of these labels during the manufacturing process to ensure the overall quality of the product.At present,the mainstream detection method still relies on visual inspection,but this method has high labor intensity,and is easily affected by subjective factors,fatigue degree and inspector experience,which leads to frequent false detection and missed detection,and the overall accuracy and efficiency are low.The defect detection method based on traditional machine vision recognizes the image by analyzing the gray level,texture and other information of the image,which has high efficiency,especially for the simple defect detection effect is good,but the complex printing plastic defect detection ability of this subject is still insufficient.With the rapid development of computer vision and artificial intelligence,especially the powerful automatic feature extraction ability of deep learning,it provides an efficient and intelligent way for industrial product defect detection.In order to improve the accuracy and efficiency of plastic label printing quality detection and defect detection,this paper studies and optimizes the defect detection method of plastic labels based on deep learning.Firstly,the image acquisition system is designed and built according to the application scene,and the test and analysis software is developed.Then,the target detection model is used to realize the identification and classification of multiple defects on plastic labels,and the detection accuracy of small targets is improved by optimizing the network structure.Finally,the DCGAN generative adversarial network is used for data enhancement to synthesize high-fidelity defect samples and further improve the accuracy of model detection.The specific research contents include:(1)An image acquisition mechanical platform is designed and customized to collect printed labels on the surface of cylindrical samples.By combining a line scan camera with a mechanical displacement/rotation platform,a specific application scenario is simulated.The hardware configuration and selection scheme of the detection system are proposed,and the overall architecture design of the software test interface is completed.Finally,the collection of sample data sets is completed.(2)The accuracy and efficiency of two-stage object detection models(R-CNN series)and single-stage detection models(YOLOv3,SSD,YOLOv5,etc.)in the detection of six common defects on plastic labels are compared and studied through experiments.It is found that YOLOv5 has the best performance(accuracy m AP is 90.91%,speed is 39.9FPS).In order to improve the performance of YOLOv5 in small object detection,the network structure is further improved.By adding a multi-scale detection head and a convolutional attention module,the detection speed is slightly reduced to 38.1FPS,but its m AP is increased to 93.61%,which is more obvious in small target defect detection,and the comprehensive performance of the model is effectively improved.(3)Due to the small number of defect samples in industrial production and the timeconsuming and labor-intensive collection process,this paper further studies the data augmentation method based on generative adversarial networks.DCGAN is used to synthesize a large number of defect samples with high fidelity,multi-scale and diversity.It is found that when the enhanced dataset of DCGAN is used to train the YOLOv5 model,the detection accuracy m AP reaches 99.6%.The experimental results of this paper also confirm the superiority of DCGAN in the fidelity,diversity and transferability of defect synthesis.Aiming at the specific requirements of the defect detection of plastic labels on the surface of cylinders,this paper has completed the hardware selection,platform customization,test software construction,and the deployment and targeted optimization of the detection algorithm based on deep learning.The research method is compatible with automated manufacturing process,and it is expected to be extended and applied to other industrial quality control and defect detection systems.