SiDIP - Deep Image Prior to Improve Siamese Network Performance on Noisy Images

Siamese networks have demonstrated remarkable capabilities in tasks like image recognition and verification. However, when they are confronted with image pairs containing a combination of clean and noisy images, their performance deteriorates notably. In this study, we investigate the effectiveness of unsupervised denoising methods, focusing particularly on Deep Image Prior and its variants, in improving the Siamese network performance on noisy-clean image pairs. Our experiments reveal a significant drop in accuracy of Siamese networks as the level of noise increases in one of the paired images. Through the application of Deep Image Prior or its variants, we aim to mitigate the adverse effects of noise and enhance the robustness of Siamese networks in evaluating image pairs comprising both clean and noisy instances. This study aims to improve the performance and accuracy of the Siamese networks on noisy images, which may be applicable to various tasks requiring similarity assessment.