Flaw Detection in White Porcelain Wine Bottles Based on Improved YOLOv4 Algorithm

Abstract

Aiming at the problems of low detection accuracy and slow detection speed in white porcelain wine bottle flaw detection, an improved flaw detection algorithm based on YOLOv4 was proposed. By adding Coordinate Attention to the backbone feature extraction network, the extracting ability of white porcelain bottle flaw features was improved. Deformable convolution is added to locate flaws more accurately, so as to improve the detection accuracy of flaws by the model. Efficient Intersection over Union was used to replace Complete Intersection over Union in YOLOv4 to improve the loss function and improve the model detection speed and accuracy. Experimental results on the surface flaw data set of white porcelain wine bottles show that the proposed algorithm can effectively detect white porcelain wine bottle flaws, the mean Average Precision of the model can reach 92.56%, and the detection speed can reach 37.17 frames/s.