Improvement of YOLOv5s-Ghost model

Abstract

Abstract Based on the YOLOv5s-Ghost model, the changes in the number of channels and feature scales of the YOLOv5s-Ghost model, as well as the structure and function of GhostConv convolution, were studied. It was found that the YOLOv5s-Ghost model has insufficient information extraction and a YOLOv5s-EGhost model with fewer model parameters and higher detection accuracy was designed. GhostConv convolution has the functions of compressing channels and downsampling. The YOLOv5s-Ghost model has a lower number of channels at low channels, and each channel doubling of the model is accompanied by a downsampling operation. Applying GhostConv convolution to the low channel of the YOLOv5s-Ghost model can cause insufficient information extraction. Therefore, ECRS and EC3Ghost2 modules were designed to be placed in the low channel of the model, and a regular 3×3 convolution was used to replace GhostConv convolution at the third downsampling point, improving the situation of insufficient information extraction in the low channel of the YOLOv5s-Ghost model. Gc3-Ghost and GSPPF modules were designed in the high channels of the model, which can reduce the number of model parameters while maintaining model accuracy. On the self-made defect dataset, the improved YOLOv5s-EGhost model reduces the number of parameters by 58.6% and 21.1% respectively compared to YOLOv5s and YOLOv5s Ghost, improves the R index by 2.0% and 1.8 percentage points respectively, and improves Map_0.5 by 0.7 percentage points and 1.8 percentage points respectively, which can meet the requirements of real-time detection.