基于深度学习的遥感旋转图像检测

Remote Sensing Rotation Image Detection Based on Deep Learning

摘要: 旋转目标检测是目前遥感图像目标检测的重要任务之一。遥感图像拥有众多小目标，检测目标密集，目标方向任意等特点。针对上述问题，提出了一种轻量级无锚框旋转目标检测算法YOLOv8-LR。首先，在主干网络和特征融合之间设计了轻量级通道空间特征增强模块，增加特征图的通道信息，提升了对遥感图像的小目标预测能力；其次，轻量化特征融合网络，在不降低检测准确率的情况下，减少模型计算复杂度；最后，增加网络的输出维度，将角度信息引入损失函数，使模型具备旋转框检测条件。实验结果表明，在DOTA数据集上，该方法检测精度高达77.2%。该模型具有良好的目标检测性能。

Abstract: Rotated object detection is currently one of the important tasks for remote sensing image object detection. Remote sensing images have characteristics such as inconsistent scales, dense detection targets, and arbitrary target directions. In response to the above issues, a lightweight anchor-free rotated object detection algorithm named YOLOv8-LR is proposed. Firstly, a lightweight channel-spatial feature enhancement module is designed between the backbone network and feature fusion to increase the channel information of the feature map and improve the prediction ability of small targets in remote sensing images. Secondly, the lightweight feature fusion network is used to reduce the computational complexity of the model without reducing the detection accuracy. Finally, the output dimension of the network is increased, and angle information is introduced into the loss function, so that the model has the conditions for rotated bounding box detection. Experimental results show that the detection accuracy of this method reaches 77.2% on the DOTA datasets, respectively. The model has good object detection performance.