改进的YOLOv4-tiny算法及其在无人机目标检测中的应用

Improved YOLOv4-tiny Algorithm and Its Application in UAV Object Detection

无人机目标检测可用于反制无人机的场景,为了便于算法在嵌入式设备部署通常需要采用轻量的模型。YOLOv4目标检测算法的轻量化版本YOLOv4-tiny具有较快的检测速度,然而其网络结构较为简单,检测精度偏低。为了进一步提升模型的检测精度,提出了YOLO-L2模型。选用YOLOv4-tiny的主干网络进行特征提取,并采用基于协调注意力机制的路径聚合网络对特征进行融合,融合过程中使用一组可学习的系数进行加权;在最深的特征层嵌入一个级联残差模块ResBlock-L2用来增大感受野并融合不同感受野特征;最后提出了边框损失函数MEIoU来替换CIoU。改进后的算法检测效果更精准,相比于YOLOv4-tiny,在VOC数据集和自制的UAV-L数据集中mAP分别提高了3.19%和3.95%,并且满足实时性的要求。

UAV object detection can be used in anti-UAV scenarios.To facilitate algorithm deployment on embedded devices,a lightweight model is often required.YOLOv4-tiny,the lightweight version of the YOLOv4 object detection algorithm,has a fast detection speed with relatively simple network structure and low detection accuracy.In order to further improve the detection accuracy,the model of YOLO-L2 is proposed.The backbone network of YOLOv4-tiny is selected for feature extraction,and a path aggregation network based on coordinate attention is used for feature fusion.In the process of fusion,a set of learnable coefficients are used for weighting.A cascade residual module named ResBlock-L2 is embedded in the deepest feature layer to enlarge receptive fields and fuse features with different receptive fields.The bounding box loss function MEIoU is proposed to replace CIoU.Compared with YOLOv4-tiny,the improved algorithm improves the mAP by 3.19% and 3.95% respectively in VOC dataset and self-made UAV-L dataset, and meets the real-time requirements.