基于改进YOLOv7的航空发动机叶片表面缺陷检测

Aircraft Engine Blade Surface Defect Detection Based on Improved YOLOv7

摘要: 航空发动机叶片表面缺陷类型复杂多样，针对目前对航空发动机叶片的检测效率和精度不高的问题，提出一种嵌入注意力机制的改进YOLOv7网络航空发动机叶片表面缺陷检测方法。通过在主干网络末尾嵌入ECA注意力机制，增强模型对于不同特征通道之间依赖关系的学习能力。用FReLU代替SiLU作为激活函数，提高边界框回归速率和目标定位精度。使用自制数据集对改进后的网络进行训练和测试。实验结果表明，YOLOv7-ECA网络在自制数据集上的精确率为97.9%，召回率为96.4%，平均精度均值为97.6%。相较于CBAM和CA注意力机制，ECA在四种缺陷的检测精度上均有一定提升。与Faster-RCNN和YOLOv5等目前的主流的目标检测模型相比，平均检测精度分别提升1.5%和2.1%，证明了此方法在航空发动机叶片缺陷检测中具有更高的检测精度。

Abstract: Aircraft engine blade surface defects are complex and diverse. To address the current challenges of low detection efficiency and accuracy, a modified YOLOv7 network with an embedded attention mechanism is proposed for aircraft engine blade surface defect detection. By embedding the ECA attention mechanism at the end of the backbone network, the model’s ability to learn the dependencies between different feature channels is enhanced. Replacing SiLU with FReLU as the activation function improves bounding box regression rate and object localization accuracy. The improved network is trained and tested on a self-developed dataset. Experimental results show that the YOLOv7-ECA network achieves a precision of 97.9%, a recall of 96.4%, and an average precision of 97.6% on the self-developed dataset. Compared with CBAM and CA attention mechanisms, ECA improves the detection accuracy of all four defect types. Compared with current mainstream object detection models such as Faster-RCNN and YOLOv5, the average detection precision increases by 1.5% and 2.1%, respectively, demonstrating the superior detection accuracy of this method for aircraft engine blade defect detection.