基于改进YOLOv4的航空发动机损伤检测方法

Aeroengine damage detection method based on improved YOLOv4

针对现有目标检测模型参数量大、检测速度慢,难以适应航空发动机孔探检测轻量化应用需求的问题,提出了基于YOLOv4目标检测算法的轻量化航空发动机损伤检测模型。设计了基于深度可分离卷积的轻量化特征融合结构,在YOLOv4的颈部结构(Neck)中,将普通卷积重构为逐通道卷积和逐点卷积的形式,有效减少了网络中的冗余参数;为进一步降低模型参数量,使用MobileNetv3作为特征提取网络。在减少参数量的同时,2种轻量化改进方法有效提高了模型的检测速度;在轻量化后的路径聚合网络(Path Aggregation Network,PANet)中加入卷积注意力模块(Convolutional Block Attention Module,CBAM),通过仅引入少量的参数来提高轻量化网络的损伤检测精度。实验结果表明,改进YOLOv4算法的平均精度均值(mean Average Precision,mAP)为89.82%,模型大小为73.29 MB,检测速度为37.3 FPS。与YOLOv4目标检测算法相比,改进YOLOv4算法以3.55%的mAP损失,使模型参数量降低了约2/3,检测速度提高了1.6倍,综合检测性能更优,可更好地满足孔探检测应用的需求,为航空发动机损伤智能化检测提供轻量化模型支撑。

The existing damage detection models have a large number of parameters and slow detection speed,which are difficult to adapt to the needs of lightweight applications of aeroengine borehole detection.A lightweight aeroengine damage detection model based on You Only Look Once version 4(YOLOv4)was proposed.By introducing depth-wise separable convolution into the neck of YOLOv4,the standard convolution was reconstructed into the channel-by-channel and point-by-point convolution,effectively reducing the redundant parameters in the network.MobileNetv3 was used as the feature extraction network to reduce the number of model parameters furtherly.Two lightweight methods effectively enhance the detection speed of the model while decreasing the number of parameters.Convolutional Block Attention Module(CBAM)was adopted into the lightweight Path Aggregation Network(PANet),which improves the detection accuracy of the lightweight network and introduces a small number of parameters.The results show that the improved model achieves a speed of 37.3 FPS,the model size is 73.29 MB,and the mean Average Precision(mAP)of damage detection reaches 89.82%.Compared with YOLOv4,the number of model parameters by about 2/3 was reduced and the detection speed was increased by 1.6 times with a 3.55%loss of mAP.The improved model has better overall detection performance,meets the needs of borehole detection applications,and provides lightweight model support for intelligent detection of aircraft engine damage.