Pill Defect Detection Based on Improved YOLOv5s Network

Abstract

To address the problems of low detection accuracy and slow speed of traditional vision in the pharmaceutical industry, a YOLOv5s-EBD defect detection algorithm: Based on YOLOv5 network, firstly, the channel attention mechanism is introduced into the network to focus the network on defects similar to the pill background, re-ducing the time-consuming scanning of invalid backgrounds; the PANet module in the network is then replaced with BiFPN for differential fusion of different features; finally, Depth-wise separable convolution is used in-stead of standard convolution to achieve the output Finally, Depth-wise separable convolution is used instead of standard convolution to achieve the output feature map requirements of standard convolution with less number of parameters and computation, and improve detection speed. the improved model is able to detect all types of defects in tablets with an accuracy of over 94% and a detection speed of 123.8 fps, which is 4.27% higher than the unimproved YOLOv5 network model with 5.2 fps.