基于注意力机制的遥感目标检测方法研究与实现

Research and Implementation of Remote Sensing Object Detection Method Based on Attention Mechanism

摘要: 遥感目标检测在环境监测、城市规划等领域具有重要应用价值，但由于遥感图像中目标尺度变化大、背景复杂以及特征提取不够鲁棒等问题，传统的检测方法难以满足实际需求。本文旨在研究一种基于注意力机制的遥感目标检测方法，以提升检测精度和鲁棒性，为遥感图像分析提供更高效的技术支持。本文提出了一种基于YOLOv8改进的遥感目标检测模型YOLO-Shuffle-MSDA，通过引入ShuffleAttention模块和多尺度特征融合模块(MSDA)来增强特征表达能力。ShuffleAttention模块通过注意力机制优化特征提取，有效区分目标与复杂背景；MSDA模块则通过多尺度特征融合，提升模型对多尺度目标的检测能力。通过多组对比实验和可视化检测结果可以看出，本文方法有效提升了检测精度，减少了误检和漏检。此外，本文设计并实现了完整的遥感目标检测系统，包括模型训练、验证和测试模块，并开发了用户友好的演示验证系统，支持单张图像、批量图像、视频以及实时摄像头检测。

Abstract: Remote sensing target detection has important application values in monitoring, urban planning and other fields, but due to the large change in target scale, complex background and insufficiently robust feature extraction in remote sensing images, detection is difficult to actually require. This paper aims to study a remote sensing object detection based on attention mechanism to improve detection accuracy and robustness and provide more efficient technical support for remote sensing images. This paper proposes a remote sensing object detection model YOLO-Shuffle-MSDA based on YOLOv8, which enhances feature expression capabilities by introducing the ShuffleAttention module and the multi-scale feature fusion module (MSDA). The ShuffleAttention module optimizes feature extraction through attention mechanism to effectively distinguish targets from complex backgrounds; the MSDA module improves the model’s detection ability of multi-scale targets through multi-scale feature fusion. Through multiple sets of comparison experiments and visual inspection results, it can be seen that this paper effectively improves detection accuracy, false detection and missed detection. In addition, this paper designs and implements a complete remote sensing object detection system, including model training, verification and testing modules, and a user-friendly demonstration verification system, supporting single images, batch images, videos and real-time camera detection.