基于全局上下文注意力的风机叶片缺陷检测

Defects Detection of Wind Turbine Blades Based on Global Context Attention

针对YOLOv3目标检测算法在风机叶片表面缺陷检测任务中存在易产生冗余框和漏检等问题,提出了一种基于全局上下文(Global Context,GC)注意力机制的风机叶片表面缺陷检测方法,简称GC-YOLOv3。该方法通过对YOLOv3的特征提取网络Darknet53中的残差块嵌入全局上下文注意力模块,使得模型能够重点关注缺陷所在位置区域,达到有效减少冗余框和降低漏检风险的目的。实验结果表明,GC-YOLOv3相比于YOLOv3准确率提升了3.38%。同时,与基于Squeezeand-Excitation和Non-local的注意力机制相比,使用全局上下文注意力机制的模型检测精度更高。

YOLOv3 object detection algorithm is prone to redundant and missed bounding boxes when detecting surface de⁃fects of wind turbine blades.This paper proposes an improved YOLOv3 algorithm for defects detection of wind turbine blades based on the global context attention mechanism,namely GC-YOLOv3.GC-YOLOv3 employs the global context attention module,and embeds it into Darknet53 feature extraction network of YOLOv3,and resulting in a model focusing more on the local areas and loca⁃tions of the defects,thereby can effectively reduce the redundant and missed bounding boxes.Experimental results show that the ac⁃curacy of GC-YOLOv3 is increased by 3.38%compared to YOLOv3.Moreover,the proposed global context attention mechanism de⁃tection algorithm can achieve the best detection accuracies compared with the attention mechanism based on Squeeze-and-Excita⁃tion and Non-local.