A Comparative Analysis Review on Panoramic Image Stitching Algorithms with Presence of Gaussian Noise

Abstract

Panoramic image stitching is a cornerstone of modern computer vision, yet its performance degrades significantly under noisy conditions particularly Additive White Gaussian Noise (AWGN). This study presents a comprehensive theoretical analysis of the robustness of three widely used feature extraction and description algorithms SIFT, SURF, and ORB in the presence of AWGN. Focusing on their behavior during feature detection, description, matching, and geometric verification. These three methods were evaluated based on three critical criteria: noise resilience, computational efficiency, and matching accuracy. The analysis reveals that SIFT demonstrates the highest theoretical robustness to Gaussian noise due to its multi-scale Difference-of-Gaussians (DoG) framework and gradient-based descriptors, which inherently suppress high-frequency noise. However, this robustness comes at the cost of high computational complexity, rendering it unsuitable for real-time applications. Conversely, ORB offers exceptional speed through binary descriptors and the FAST detector but exhibits marked sensitivity to intensity perturbations caused by AWGN, leading to significant performance degradation in noisy environments. SURF emerges as the optimal compromise: leveraging integral images, it achieves near-SIFT-level robustness while maintaining significantly lower computational demands. Research underscores the importance of the Fast Library for Approximate Nearest Neighbors (FLANN) library in enhancing feature matching and the Random Sample Consensus (RANSAC) algorithm, offering a methodological framework for choosing the appropriate algorithm in panoramic compositing systems. It emphasizes balancing noise characteristics, speed, and accuracy requirements