联合结构重参数和YOLOv5的航拍红外目标检测

Infrared Aerial Object Detection Based on Structural Re-Parameterization and YOLOv5

无人机进行红外航拍目标检测在交通、农业和军事等方面有着广泛应用。该领域的主要挑战有目标较小、相互遮挡、非刚体形变大以及红外成像纹理信息少、边缘特征弱等。针对以上问题,基于YOLOv5和结构重参数化优化思想,提出了一种针对航拍场景的目标检测模型Rep-YOLO。首先,在主干网络中引入RepVGG模块,提升模型特征提取能力;在模型推理时对RepVGG模块的多分支进行结构重参数化,减少网络分支和结构复杂度。其次,结合数据特征,改进检测网络颈部的路径聚合网络,提升检测算法在机载平台的精度-速度均衡能力。最后,在两个公开红外数据集进行对比实验,表明该算法的有效性。以南航ComNet航拍数据集为例,统计结果显示主要检测指标各类平均精度(mean Average Precision,mAP)提升5.9%,同时参数量和模型大小分别减少约29.7%和23.2%。另外,对Rep-YOLO在典型机载平台Jetson Nano上进行了模型部署验证,为航拍场景的检测算法改进和实际应用提供了可靠的技术支撑。

Infrared aerial object detection has been widely used in transportation,agriculture,military security,and other areas.The main challenges are small objects,mutual occlusion,little texture information,weak edge features,and large deformation of non-rigid bodies.To address these problems,based on YOLOv5 and structural Re-Parameterization(Rep),an improved object detection network Rep-YOLO is proposed for infrared aerial object detection.Firstly,the RepVGG module is introduced in the backbone network to improve the model feature extraction capability.During the model inference,the branches of the RepVGG module are structurally re-parameterized to reduce the branch and the complexity of the network structure.Secondly,the path aggregation network(PANet)in the neck of the detection network is improved by combining the priori feature,to increase the accuracy and speed balance capability.Finally,experiments are conducted on two publicly available infrared datasets,showing that the algorithm can effectively detect aerial infrared objects.Compared with the baseline(YOLOv5s),the statistical results on ComNet dataset show the mean Average Precision(mAP)is increased by 5.9%,while the parameters and model size are reduced by about 29.7%and 23.2%,respectively.In addition,the model deployment verification of our Rep-YOLO is carried out on the airborne platform Jetson Nano.It provides reliable technical support for the improvement of the detection algorithm and its practical application with UAV platforms.