An industrial interference-resistant gear defect detection method through improved YOLOv5 network using attention mechanism and feature fusion

Abstract

Automatically detecting the metallic gear defects may encounter the misdetected and undetected errors when the defect features are very close to the interferences of oil stains and image shadow. To address these challenges, in this paper an industrial interference-resistant gear surface defect detection method is presented based on the improved YOLOv5 network. The method combines the approach using the attention module CBAMC3 constituted by CBAM and C3 modules together with the analysis of the module BiFPN_concat, which makes the network capable of extracting and fusing defect features, respectively. The cosine annealing function is then employed to improve the learning ability of the network by modifying the learning rate. The experimental results indicate that the improved YOLOv5 network enhances both the recall rate by 13.1% and the mAP@0.5 by 12% compared with the original YOLOv5 network, and also outweighs the classical algorithms in detection speed with average 25 frames per second.