面向无人机航拍图像的多尺度目标检测研究

Multiscale Target Detection for UAV Aerial Images

摘要:

针对无人机航拍图像背景复杂、小目标占比高且分布不均导致的现有算法精度不佳等问题,提出了一种面向无人机航拍图像的多尺度目标检测网络VTO-YOLOv8。首先,采用WIoU v3作为边界框回归损失函数,并使用明智的梯度分配策略,这一策略将使网络更加关注普通质量样本,从而提高其定位能力；其次,设计四层T-BiFPN结构,加强浅层特征和深层特征的融合；此外,设计C2f-DBB多分支模块,在不增加计算量的前提下,提升检测性能；同时,使用聚焦调制模块,加强不同尺度信息的交互。实验结果表明,网络在Visdrone2019数据集上相较基准模型在mAP50和mAP指标上分别提高了9.0%和5.9%,同时参数降低了22.6%,可更好地应用于无人机航拍目标检测中。

Abstract:

A multiscale target detection network, VTO-YOLOv8, for unmanned aerial vehicle (UAV) images is proposed to address the low accuracy of existing algorithms caused by complex backgrounds, a high proportion of small targets, and uneven distributions. First, wise intersection over union (WIoU) v3 was used as the bounding-box regression loss, and a wise gradient allocation strategy was employed for the network to focus more on regular quality samples and improve localization ability. Second, a four-layer target bi-directional feature pyramid network (T-BiFPN) structure was designed to strengthen the integration of shallow and deep features. Furthermore, a faster implementation of CSP bottleneck with diverse branch blocks (C2f-DBB) module was designed to improve the detection performance of the network without increasing computational complexity. In addition, a focal modulation module was used to enhance the interaction of information at different scales. The experimental results demonstrated that the proposed network improved the mean average precision (mAP) and mAP50 by 5.9% and 9.0%, respectively, compared with those of the baseline network on the Visdrone2019 dataset. Moreover, the network parameters were reduced by 22.6%. The proposed method can be applied to target detection in UAV aerial photography.