基于改进YOLOX-s的风机叶片表面缺陷检测

Surface Defect Detection of Wind Turbine Blades Based on Improved YOLOX-s

提出一种改进YOLOX-s的缺陷检测方法。主要工作包括以下3个方面:第一,通过Imgaug数据增强策略重新构建了风机叶片缺陷数据集,弥补真实场景下的数据量不足;第二,采用模型压缩策略,对骨干网络的部分模块进行删减,并引入深度可分离卷积,提升模型的推理速度,重构CBS卷积块为DSCBM模块,用于稳定网络性能;第三,引入GiraffeNeck融合机制和CA坐标注意力机制,提高模型对不同尺度特征的融合能力以及对缺陷目标的检测能力,对Head层进行改进,删减部分冗余的卷积块,进一步提升检测速度。实验结果表明,与YOLOX-s模型相比,mAP值提升2.6%,检测速度提高39帧/s。

In this paper,an improved defect detection method of YOLOX-s is proposed,the main work includes three aspects.First,the fan blade defect data set is reconstructed through Imgaug data enhancement strategy to make up for the insufficient data volume in real scenarios.Second,the model compression strategy is adopted to delete some modules of the backbone network and introduce deep separable convolution to improve the reasoning speed of the model,reconstructed CBS convolutional block as DSCBM module to stabilize network performance.Third,GiraffeNeck fusion mechanism and CA coordinate attention mechanism are introduced to improve the fusion ability of different scale features and the detection ability of defect targets,the Head layer is improved,some redundant convolutional blocks are deleted,and the detection speed is further improved.The experimental results show that compared with the YOLOX-s model,the mAP value is increased by 2.6%and the detection speed is increased by 39 frames/s.