基于改进YOLOv5的遥感图像目标检测研究

Research on Object Detection in Remote Sensing Images Based on YOLOv5

摘要:

针对遥感图像中背景复杂度高、目标尺寸多样所导致的目标检测精度低的问题,提出一种基于改进 YOLOv5的遥感图像目标检测算法。该算法将具有Transformer风格的ConvNeXt网络作为主干网络,以克服卷积神经网络(CNN)结构的局限性,捕获更多全局信息。引入 SimAM 注意力机制在不增加网络参数的情况下,推断出特征图的3D注意力权值,提高网络的稳定性以及抗干扰能力。同时采用全局显式集中调节方案的集中特征金字塔(CFP),捕获全局长距离依赖关系以及遥感图像的局部关键区域信息。将本文提出的算法在 RSOD 数据集上进行消融实验,结果表明,本文提出的算法能够显著提高遥感图像目标检测的平均准确率。

Abstract:

To address the problem of low detection accuracy caused by high background complexity, diverse target sizes, and an abundance of small objects in remote sensing images, a remote sensing image object detection algorithm based on improved YOLOv5 is proposed. This algorithm employs a ConvNeXt network with Transformer-style architecture as the backbone network to overcome the limitations of CNN structures and capture more global information. The SimAM attention mechanism is introduced to infer 3D attention weights of the feature maps without increasing the network parameters, thereby enhancing the stability and anti-interference capability of the network. Additionally, a concentrated feature pyramid (CFP) with a global explicit focus adjustment scheme is utilized to capture long-range global dependencies and local key region information in remote sensing images. The proposed algorithm is evaluated on the RSOD dataset through ablation experiments, and the results demonstrate a significant improvement in the average accuracy of remote sensing image object detection.