基于注意力机制与特征融合的航拍图像小目标检测算法

Small Object Detection Algorithm in Aerial Images Based on Attention Mechanism and Feature Fusion

针对无人机航拍图像背景复杂、目标密集且小目标较多等问题,提出一种基于注意力机制与特征融合的航拍图像小目标检测算法。首先在主干网络部分嵌入多头自注意力机制(MHSA),有效整合全局特征信息;然后添加小目标检测层与BiFPN结构,避免检测小目标时尺度不一导致语义丢失,加深浅层语义与深层语义结合,并在特征金字塔网络输出端后引入自适应空间特征融合(ASFF)模块,提高特征尺度不变性,改善不同尺度特征图的检测精度;其次在特征融合网络中添加CA注意力机制,避免背景信息的干扰;最后使用损失函数MPDIoU,提高收敛速度和回归精度。在VisDrone2019-DET数据集上进行实验,与YOLOv5s相比,改进后算法的mAP提高了7.8%,对小目标检测任务具有较好的检测效果。

A small object detection algorithm based on attention mechanism and feature fusion is proposed to solve the problems of complex background,dense targets and many small targets in UAV aerial images.Firstly,the multi-head self-attention mechanism(MHSA)is embedded in the backbone network to effectively integrate the global feature information.Then,the small target detection layer and BiFPN structure are added to avoid semantic loss caused by different scales when detecting small targets,and deepen the combination of shallow semantics and deep semantics.After the output of the feature pyramid network,the adaptive spatial feature fusion(ASFF)module is introduced to improve the feature scale invariance and improve the detection accuracy of feature maps at different scales.Secondly,the CA attention mechanism is added to the feature fusion network to avoid the interference of background information.Finally,the loss function MPDIoU is used to improve the convergence speed and regression accuracy.Experiments were carried out on VisDrone2019-DET dataset,and compared with YOLOv5s,the mAP of the improved algorithm increased by 7.8%,which has a good detection effect on small target detection tasks.