基于自适应空间融合的遥感图像目标检测器

A Remote Sensing Image Object Detector Based on Adaptive Spatial Fusion

摘要: 随着遥感技术的蓬勃发展，日益庞大的遥感数据使得传统的图像处理手段已经不足以满足需求，因此，将深度学习目标检测技术应用于遥感领域成为了最优解决方案。然而，由于遥感图像小目标排列密集、尺度变化剧烈，使得直接将适用于自然场景的目标检测算法迁移到遥感领域的效果不佳。本文选择YOLOX为基础网络进行改进，在特征提取网络后加入自适应空间特征融合结构，将深层特征信息与浅层特征信息融合，提升小目标的识别率。此外，本研究对样本标签匹配策略进行了优化，以解决高长宽比类目标物体角度偏移敏感的问题，并通过修改损失函数降低正负样本不平衡问题带来的影响。我们在遥感图像数据集DOTA上进行训练和测试，实验结果表明，改进的YOLOX算法检测效果更好，mAP达到了79.07%，比YOLOX提高了2.75%。另外，在HRSC2016数据集上也进行了实验，实验证明了模型具有优秀的鲁棒性。

Abstract: With the rapid development of remote sensing technology, the increasingly large remote sensing data has rendered traditional image processing techniques inadequate to meet the demands. As a result, applying deep learning object detection techniques to the remote sensing field has become the optimal solution. However, the dense arrangement of small objects and drastic scale variations in remote sensing images make the direct migration of object detection algorithms applicable to natural scenes to the field of remote sensing ineffective. In this paper, we improve the YOLOX algorithm by incorporating an adaptive spatial feature fusion structure after the feature extraction network, which fuses deep and shallow features to enhance the recognition rate of small objects. In addition, this study optimized the sample label assignment strategy to address the issue of angle offset sensitivity for narrow class target objects with high aspect ratios. Furthermore, the impact of the imbalance between positive and negative samples was mitigated by modifying the loss function. We conduct experiments on the DOTA remote sensing image dataset for training and testing. The results demonstrate that the improved YOLOX algorithm achieves better detection performance, with a mAP of 79.07%, which is 2.75% higher than that of YOLOX. Furthermore, we conduct experiments on the HRSC2016 dataset, which confirms the model’s excellent robustness.