Zeng Qinyan, Hui Bin, Liu Zhaoji, Xu Zheng, He Miao
Key Laboratory of Opto-Electronic Information Processing, Chinese Academy of Sciences, Shenyang 110016, China.
Shenyang Institute of Automation, Chinese Academy of Sciences, Shenyang 110016, China.
Sensors (Basel). 2025 Feb 22;25(5):1345. doi: 10.3390/s25051345.
Multi-temporal remote sensing image matching is crucial for tasks such as drone positioning under satellite-denial conditions, natural disaster monitoring, and land-cover-change detection. However, the significant differences between multi-temporal images often lead to the reduced accuracy or even failure of most image matching methods in these scenarios. To address this challenge, this paper introduces a Discrete Cosine Transform (DCT) for frequency analysis tailored to the characteristics of remote sensing images, and proposes a network that combines the DCT with attention mechanisms for multi-scale feature matching. First, DCT-enhanced channel attention is embedded in the multi-scale feature extraction module to capture richer ground object information. Second, in coarse-scale feature matching, DCT-guided sparse attention is proposed for feature enhancement, which suppresses the impact of temporal differences on matching while making the amount of computation controllable. The coarse-scale matching results are further refined in the fine-scale feature map to obtain the final matches. Our method achieved correct keypoint percentages of 81.92% and 88.48%, with average corner errors of 4.27 and 2.98 pixels on the DSIFN dataset and LEVIR-CD dataset, respectively, while maintaining a high inference speed. The experimental results demonstrate that our method outperformed the state-of-art methods in terms of both robustness and efficiency in the multi-temporal scenarios.
多时相遥感图像匹配对于卫星信号受阻情况下的无人机定位、自然灾害监测和土地覆盖变化检测等任务至关重要。然而，多时相图像之间的显著差异常常导致大多数图像匹配方法在这些场景中的准确性降低甚至失败。为应对这一挑战，本文引入了一种针对遥感图像特征进行频率分析的离散余弦变换（DCT），并提出了一种将DCT与注意力机制相结合用于多尺度特征匹配的网络。首先，将DCT增强通道注意力嵌入到多尺度特征提取模块中，以捕获更丰富的地物信息。其次，在粗尺度特征匹配中，提出了DCT引导的稀疏注意力用于特征增强，在控制计算量的同时抑制时间差异对匹配的影响。粗尺度匹配结果在细尺度特征图中进一步细化以获得最终匹配。我们的方法在DSIFN数据集和LEVIR-CD数据集上分别实现了81.92%和88.48%的正确关键点百分比，平均角点误差分别为4.27和2.98像素，同时保持了较高的推理速度。实验结果表明，我们的方法在多时相场景的鲁棒性和效率方面均优于现有方法。
