融合多尺度和上下文的无人机救援小目标检测算法

Fusion of multi-scale and context for small target detection algorithm of unmanned aerial vehicle rescue

针对无人机(UAV,unmanned aerial vehicle)图像中小目标所包含的特征信息少,导致模型检测精度不足的问题,面向无人机海面救援任务提出了一种融合多尺度和上下文信息的图像小目标检测算法。首先,针对小目标特征信息设计上下文增强模块,通过增强特征层的上下文信息,有效地增加了模型对小目标的处理能力。其次,为提高模型的鲁棒性,设计了空间注意力模块加强对重要特征的学习。最后,使用平衡L1损失函数优化基线算法的损失函数,加强了模型检测时的稳定性。基于Tiny-Person数据集,与基准算法进行大量实验对比,所提算法在AP50_tiny上提高了2.06%,一定程度上提高了对海面小目标的检测性能,对救援行动具有积极影响。

Aiming at the problem of insufficient feature information contained in small targets under unmanned aerial vehicle(UAV)images that led to insufficient detection accuracy of the model,a small target detection algorithm for UAV sea rescue images that integrated multi-scale and contextual information was proposed.Firstly,context enhancement module was designed for small target feature information,which effectively enhanced the ability of the model to process small targets by enhancing the contextual information of the feature layer.Secondly,to improve the robustness of the model,spatial attention module was designed to enhance the learning of important features.Finally,balance L1 loss was used to optimize the loss function of the baseline algorithm and enhance the stability of the model during the process of detection.Based on the Tiny-Person dataset,through extensive experimental comparison with the benchmark algorithm,the proposed algorithm improves the detection performance of small targets on the sea surface by 2.06%on AP50_tiny,which has a positive impact on rescue operations.