多尺度特征与注意力检测头的轻量化FOD检测

Lightweight FOD Detection with Multi-Scale Features and Attention Detection Heads

针对现阶段机场跑道异物碎片(FOD)目标检测算法存在的缺陷,进行了降低参数量和提高精度的改进。以你只看一次(YOLO)v5s目标检测算法为基础,提出多尺度特征与注意力检测头的轻量化FOD检测算法。首先,提出一种全新的轻量化网络结构。该结构使用深度可分离卷积和逐点卷积,并设计大卷积核架构,使模型感受野提升,从而解决大量特征图冗余问题。接着,融合多尺度特征图。通过移除大目标检测层、增加小目标检测层,在提升小目标检测能力的同时降低网络参数量。最后,提出一种动态头部框架来统一目标检测头和注意力,通过连贯地结合多个自注意力机制,进一步提升了网络检测精度。试验结果表明:所提出的使用鬼影卷积大卷积核架构下的多尺度特征注意力检测头的YOLOv5s(GRD-YOLOv5s)网络的参数量减少为3.39 MB,仅为原网络的48%;平均检测精度从98.40%提升至99.45%;检测速度为53.42帧/秒。该网络的提出为实现对小目标的准确检测提供了新思路。

Aiming at the defects of the foreign object debris(FOD)target detection algorithm of airport runway at the present stage,the improvement of reducing the number of parameters and increasing the accuracy is carried out.Based on you only look once(YOLO)v5s target detection algorithm,a multi-scale feature and attention detection heads lightweight FOD detection algorithm is proposed.Firstly,a new lightweight network structure is proposed.The structure uses depth-separable convolution and point-bypoint convolution,and designs a large convolutional kernel architecture to enhance the model sensory field,thus solving the problem of redundancy of many feature maps.Then,the multi-scale feature maps are fused.The number of network parameters is reduced by removing the large target detection layer and adding the small target detection layer,while improving the small target detection capability.Finally,a dynamic head framework is proposed to unify the target detection head and attention,which further improves the network detection accuracy by coherently combining multiple self-attention mechanisms.The experimental results show that the proposed Ghost RepLKNet Dyhead YOLOv5s(GRD-YOLOv5s)network parameter quantity reduced to 3.39 MB,which is only 48%of the original network;the average detection accuracy is improved from 98.40%to 99.45%;the detection speed is 53.42 frames/s.The proposed network provides a new idea to realize the accurate detection of small targets.