基于可变形卷积与注意力的无人机航拍车辆目标检测算法

Deformable ConvNets and attention based object detection algorithmfor vehicles in UAV aerial photography

在无人机航拍图像中,车辆目标较小,尺度变化大,背景复杂且分布密集,导致精度过低的问题。因此,提出一种基于改进的YOLOv5的无人机航拍图像车辆目标检测算法。增加小目标检测层,减少小目标特征丢失,从而提高小目标检测精度;设计了一个名为DAC的新特征提取模块,它融合了标准卷积、可变形卷积和通道空间注意力机制,旨在增强模型对车辆尺度变化的感知能力,并让模型聚焦于复杂背景下的车辆目标;将损失函数更改为Focal-EIoU,以加速模型收敛速度,同时提高小目标车辆的检测精度。使用Soft-NMS代替YOLOv5中采用的非极大值抑制,从而改善目标密集场景下的漏检和误检情况。在VisDrone2019数据集上进行了消融实验、对比实验和结果可视化。改进后的模型平均精度(mAP)比基线模型提高了8.4%,参数量和GFLOPs仅增加了4.8%和3.79%,验证了改进策略的有效性和优越性。

In UAV aerial images,the vehicles(the objects)are small,the scale changes greatly,and the background is complex and distributed densely,which results in low accuracy.Therefore,an improved YOLOv5 based object detection algorithm for vehicles in UAV aerial images is proposed.A small object detection layer is added to reduce the feature loss of small objects,so as to improve the accuracy of small object detection.A new feature extraction module called DAC,which combines standard convolution,deformable ConvNet(DCN)and channel space attention mechanism,is designed,which aims to enhance the model's perception of changes in vehicle scale and allow the model to focus on vehicles(the objects)under complex backgrounds.The loss function is changed to Focal⁃EIoU to speed up the convergence of the model and improve the detection accuracy of small vehicles(the objects).The Soft⁃NMS is used to replace the non⁃maximum suppression used in YOLOv5,so as to improve missed detections and false detections in scenarios with dense objects.Ablation experiments,comparison experiments and result visualization are conducted on the VisDrone2019 data set.The mean average precision(mAP)of the improved model is 8.4%higher than that of the baseline model,and its number of parameters and GFLOPs are only increased by 4.8%and 3.79%.The effectiveness and superiority of the improved strategy are verified.