Date of Award
5-30-2025
Document Type
Thesis - SCU Access Only
Publisher
Santa Clara : Santa Clara University, 2025
Degree Name
Master of Science (MS)
Department
Computer Science and Engineering
First Advisor
Nam Ling
Abstract
Image denoising is a fundamental low-level vision task, crucial for enhancing image quality and ensuring reliable high-level visual analysis. However, existing discrete generative prior-based methods often require separate codebooks for specific image categories (e.g., faces, buildings), limiting their generalization to diverse real-world noise. To address this, we propose AdaDenoise, a class-agnostic image denoising framework based on adaptive codebook fusion. By dynamically learning a weight map from the input image, AdaDenoise selectively combines a set of base codebooks to construct a customized prior tailored to the image content. This adaptive mechanism allows the model to flexibly adjust its latent representation and improve robustness against unknown noise patterns. Experimental results on the CBSD68 dataset demonstrate that AdaDenoise achieves competitive performance, particularly excelling in preserving fine details and generalizing across varied noisy image domains such as natural scenes and textured surfaces.
Recommended Citation
Xu, Xiaoya, "Learning Image-Adaptive Codebooks for Class-Agnostic Image Denoising" (2025). Computer Science and Engineering Master's Theses. 53.
https://scholarcommons.scu.edu/cseng_mstr/53
