LTIRM: learnable transformation for illumination-robust matching

Image matching is a fundamental task in computer vision, yet its performance often degrades significantly under varying illumination conditions. This paper focuses on addressing the problem of illumination domain adaptation in image matching to improve the generalization capability and robustness of models in the target illumination domain. We propose a novel framework named Learnable Transformation for Illumination-Robust Matching (LTIRM). A global transformation module is introduced to automatically generate training samples with enhanced domain diversity by applying parameterized adjustments to the brightness, contrast, saturation, and hue of source domain images. A domain classifier is proposed to maximize the domain discrepancy among images from different domains. In addition, a feature distribution alignment loss is incorporated to explicitly minimize the distribution gap between feature maps from different domains in the latent space, thereby facilitating the learning of illumination-invariant essential features. Experiments demonstrate that the proposed LTIRM framework outperforms several state-of-the-art sparse image matching methods on the dataset SEUIAM.