Wine Box Product Label Defect Detection System Based On Deep Learnin

The label is an important identification of the product,and its printing quality will directly affect the overall appearance of the product.At present,most of the detection of label defects is manual visual inspection,which has the disadvantages of low efficiency and high false detection rate.A small part is Image processing detection has high environmental requirements,complex debugging,and poor stability.Therefore,it is of great practical significance to design a portable and efficient label defect detection system.The label defect detection system based on deep learning can learn the characteristics of label defects and establish a defect recognition model through a large amount of training data.In this thesis,the relevant detection algorithm in the field of deep learning is applied to the detection of product label defects,an improved label defect detection algorithm based on YOLOv4 is studied,and a corresponding detection system is built.Taking the defect of wine box label as the research object,and focusing on the detection of wine box label defect,the main work content of this thesis is as follows:(1)Establish a data set for the defects of wine box labels,and expand the data set in many different ways.Aiming at the development of target detection algorithms in deep learning and the goals studied in this thesis,the self-made wine box label flaw dataset is used to test different target detection algorithms,including Faster R-CNN,SSD,and YOLOv4.After analyzing and comparing the experimental results,and testing requirements,the YOLOv4 algorithm with average performance was selected for improvement and optimization.(2)Aiming at the problems of low accuracy,slow reasoning speed and large model size in the label defect detection method based on YOLOv4,an improved label defect detection method based on YOLOv4 is proposed.First,replace the YOLOv4 backbone network with Ghost Net,which greatly reduces the parameter amount of the backbone feature extraction network and reduces the model size.Secondly,an attention mechanism is added to the backbone network to enhance feature extraction capabilities and improve accuracy without increasing computational complexity.Then use the blueprint convolution to reduce the computational complexity of the algorithm while improving the detection accuracy to achieve lightweight,and finally use an activation function with better performance.Through the experiment of the selfmade label defect sample set,the actual measurement results show that the proposed improved algorithm is lightweight and efficient.Compared with the original algorithm,the m AP accuracy and detection speed are improved,and the model size is reduced,which can solve the existing problems.(3)According to the improved YOLOv4 label defect detection algorithm proposed in this thesis and the actual needs in product production,a set of label defect detection system based on Qt Designer,Py Qt5,MYSQL is designed,including administrator and employee login,registration,leaflet Image detection,information statistics,etc.,realize the demand for product surface defect detection in product production.The system has been used in Bozhou Fuda Printing Co.,Ltd.and has achieved good results.