Defect Detection Method of Wind Turbine Blades Based on Improved YOLOv4

Abstract

A lightweight defect detection model for wind turbine blades is needed to meet the application in mobile devices and embedded devices. Though there are many kinds of research on Image Detection, designing a robust and effective defect detection model is still an open issue. Therefore, this paper proposes a lightweight target detection algorithm based on the regression-based YOLOv4 by simplifying the backbone network, pruning the model with channel attention, and simplifying the anchor box. From the perspective of backbone network simplification, we designed a novel framework named Tiny-GhostNet to replace the original CSPDarknet53 network. Channel attention-based model pruning mainly utilizes channel attention to remove those unimportant channels. The simplification of anchor boxes aims to simplify predefined anchor box settings and density distribution.