基于改进YOLOv5s的遥感影像小目标检测

Small object detection in remote sensing images based on improved YOLOv5s

针对现有目标检测模型难以准确检测小目标并存在大量误检与漏检的问题,该文提出了一种基于YOLOv5s的遥感影像小目标检测算法。首先,在主干网络加入SimAM无参数注意力机制模块,在不增加额外参数的情况下,使算法更加关注重要特征,同时抑制遥感影像中背景信息;然后,采用SPD-Conv模块进行下采样,避免了特征信息的丢失;之后,引入了四尺度检测特征融合网络,以获得更丰富的多尺度特征信息,并优化融合网络和输出检测头结构。最终实现对于遥感影像小目标的精准检测。在自制的LEVIR-4SC遥感影像小目标数据集上的实验结果表明,与YOLOv5s相比,提出算法的模型参数减少了6.3%,精确度P和平均精度均值mAP达到了90.7%和83.7%,提升了6.5%和4.2%。与YOLOv8等经典方法相比参数量更少,模型更轻量化,在精确度、平均精度方面均有提升,证明提出算法具有更好的检测效果。

Due to the difficulty of accurately detecting small objects in existing object detection models and the existence of a large number of false positives and false negatives, this paper proposes a small object detection algorithm for remote sensing images based on YOLOv5s. Firstly, a SimAM parameter-free attention mechanism module is added to the backbone network to make the algorithm pay more attention to important features without adding additional parameters, while suppressing background information in remote sensing images. Then, SPD-Conv modules are used for down-sampling to avoid the loss of feature information. Subsequently, a four-scale detection feature fusion network is introduced to obtain richer multi-scale feature information and optimize the fusion network and output detection head structure. Finally, accurate detection of small objects in remote sensing images is achieved. The experimental results on the self-made LEVIR-4SC remote sensing image small object dataset show that compared with YOLOv5s, the model parameters of the proposed algorithm are reduced by 6.3%,and the accuracy(P) and average accuracy mean(mAP) reach 90.7% and 83.7%,which are improved by 6.5% and 4.2%. Compared with classic methods such as YOLOv8,the proposed algorithm has fewer parameters and a lighter model, which improves both accuracy and average accuracy, proving that the proposed algorithm has better detection performance.