Performance evaluation of state-of-the-art multimodal remote sensing image matching methods in the presence of noise

To date, various image registration approaches have been conducted to deal with distortions between multimodal image pairs. However, significant existing noise as an unavoidable issue deteriorates many conventional and advanced methods. The critical key is to choose a highly robust local feature detection and description method as the principle for many matching frameworks. However, few studies have concentrated on dealing with the noise issue. For this purpose, this paper evaluates the most well-known and state-of-the-art feature descriptors against artificial sequential noise levels. The employed methods consist of various handcrafted learning-based descriptors. It is further indicated that in addition to the designed structural feature map, multiple criteria, such as spatial arrangement, and the magnitude of the support area, play roles in achieving successful matching, especially in the presence of dramatic noise and complex distortion between multimodal images. Moreover, to filter out the noisy features, the employed local feature detectors are integrated with the uniform competency algorithm. Experimental results demonstrate the overall superiority (20.0% on average) of the MKD (multiple-kernel descriptor) due to advanced designed integrated kernels and polar arrangements.