A gray box explainable deep learning approach for blind deblurring of motion blurred images
Date of Publication
4-2024
Document Type
Master's Thesis
Degree Name
Master of Science in Electronics and Communications Engineering
Subject Categories
Electrical and Computer Engineering | Engineering
College
Gokongwei College of Engineering
Department/Unit
Electronics And Communications Engg
Thesis Advisor
Lawrence Materum
Defense Panel Chair
Edwin Sybingco
Defense Panel Member
John Anthony Jose
Joel Ilao
Abstract/Summary
In deblurring technologies, the lack of interpretability limits the ability to debug, troubleshoot, and improve the method effectively, hence the need for enhancing explainability has become increasingly important. This thesis addresses the challenge of developing a method that enhances the explainability of deep learning models. The objective is to strike a balance between performance and explainability by combining the power of deep learning techniques with meaningful explanations for the deblurring process. The proposed methodology involves the design of a gray box explainable deep-learning network that involves blur and sharp features estimation, blur kernel estimation, and latent image estimation that induces explainability. The best-performing model achieves 26 dB, 25 dB, 35 dB, and 28 dB accuracy levels in the GoPro, HIDE, RealBlur-R, and RealBlur-J datasets, respectively, reaching an average accuracy improvement of 1 dB. Models emphasizing explainability prioritize transparent operations for easier interpretation but may sacrifice visual fidelity and efficiency. Conversely, models focused on accuracy often employ complex architectures for better performance, potentially reducing interpretability and increasing computational costs. By achieving a satisfactory balance between accuracy and explainability, this work aims to unlock the potential for advanced deblurring technologies that produce visually appealing results while providing transparent and understandable insights into the deblurring process.
Abstract Format
html
Language
English
Format
Electronic
Keywords
Image processing; Machine learning
Recommended Citation
Barlis, M. B. (2024). A gray box explainable deep learning approach for blind deblurring of motion blurred images. Retrieved from https://animorepository.dlsu.edu.ph/etdm_ece/31
Upload Full Text
wf_yes
Embargo Period
4-4-2024
