Comparative Evaluation of Deep Learning and Traditional Methods for Retinal Image Stitching on Mobile-Based Imaging Devices

This paper presents a comparative study of traditional feature-based, area-based, and deep learning-based methods for retinal image stitching with a focus on lowresolution, smartphone-attained fundus images. Traditional methods such as SURF and MODE-based registration and mutual information-based template matching were compared with deep convolutional neural networks specifically designed for homography estimation. Experiments on multiple public and semi-clinical data sets reveal that while the traditional approaches are still reliable in controlled imaging scenarios, deep learning approaches are always better regarding stitching accuracy, robustness against noise, and ability to adapt to diversified imaging scenarios. Particularly, lean deep models offer superior PSNR and structural quality, corroborating their suitability for real-time mobile teleophthalmology tasks. These findings show the greater potential of deep learning for enhancing portable ocular diagnostics and opening new opportunities for future development in real-time panoramic reconstruction and hybrid registration architecture.