【摘要】 通过人工对飞机导管喇叭口(AFT)缺陷进行检测过程中存在一定误差且检测率低下。为减少人工检测带来的误检、漏检及提高检测工效，提出了一种基于YOLOv4改进模型的检测飞机导管常见缺陷的方法。通过聚类分析调整锚盒的大小，以更好地匹配小目标和复杂结构的特征。在骨干特征提取网络和空间金字塔池结构输出的不同特征层之后添加卷积层，以提高网络复杂度换取网络对缺陷特征的提取能力。实验结果表明，改进模型在对飞机导管喇叭口缺陷检测中的AP值为92.21%,比原始YOLOv4算法提高16.67%,对单个图像的平均检测时间算法运行速度与原始算法相较没有明显变化。更多还原

【Abstract】 There are problems of accuracy and efficiency in the manually detection of AFT detects.To reduce false or missed detections caused by manual inspection and improve inspection efficiency, this paper proposes a method for detecting common defects in AFT based on the YOLOv4 improved model.Firstly, adjust the size of the anchor box through clustering analysis to match the features of small targets and complex structures better.Then, after different feature layers produced by backbone feature extraction network and spatial pyramid pool structure, this paper adds the convolution layer to these feature layers to improve the network′s ability to extract defect features.These improvements result that the model has an AP value of 92.21% in AFT defectsdetection, which is 16.67% higher than the original YOLOv4 algorithm.But the average detection time of a single image has not significant increasing compared to the original algorithm.更多还原