Abstract: This paper presents a fast and robust image matching approach based on the cosine M-estimator kernel and the Fast Fourier Transform (FFT). We show that the robust cosine M-estimator can be effectively used to compare image intensities through correlation of transformed images. The speed of the method derives from the use of FFT to compute correlation. Its computational complexity is O (N log N), compared to O(N²) for direct matching. Experimental results demonstrate that the proposed method maintains high matching accuracy even in the presence of up to 60% outliers and 60% occlusion. It outperforms traditional approaches such as Sum of Squared Differences (SSD) and Normalized Cross-Correlation (NCC). Moreover, the proposed approach achieves superior performance compared to recent deep learning-based methods such as LoFTR. By combining the cosine M-estimator with FFT, the method is suitable for real-time applications that require robust matching under challenging conditions.

Keywords: Image Matching, Robust Correlation, Cosine M-estimator, Robust Statistics, FFT, Fast Algorithms

Pages: 287-291