Robust Day-Night Image Matching Across Extreme Illumination Variations: A Comparative Study of Deep Learning and Classical Methods

Abstract

The clinical applicability of Magnetic Resonance Imaging (MRI) is often constrained by prolonged scanning durations, particularly in emergency scenarios. To address this, this study conducts a comparative analysis between a traditional greedy algorithm, Orthogonal Matching Pursuit (OMP), and a deep learning alternative known as Generative Adversarial Networks for compressed sensing (GANCS). Empirical evidence suggests that GANCS outperforms the classical method by delivering higher fidelity images from highly undersampled data, while also ensuring rapid processing speeds following the training phase. obust image matching across extreme illumination conditions remains a critical challenge in computer vision, paricularly for sustainability-related applications such as smart city monitoring, urban safety, and environmental surveillance. This study evaluates the performance of local feature matching techniques under severe day–night illumination variations using images from the AMOS dataset. I compare a deep learning–based approach, MatchNet, with a classical template-matching method based on distance transforms and normalized cross-correlation. Experimental results demonstrate that while deep learning models exhibit strong feature repeatability, they struggle to achieve reliable matching precision under extreme illumination changes. In contrast, the classical approach shows superior mean average precision, particularly for cross-domain image retrieval tasks. The findings highlight the importance of task-specific modeling choices and suggest directions for improving illumination robust feature matching in sustainable vision systems