Comparison Of Blurred Image Restoration Methods Using CNN, Non-Local Means (NLM), and Lucy-Richardson

Abstract

Purpose: Blurred images caused by camera motion, poor lighting, or inaccurate focus are common challenges in digital imaging. These issues not only affect visual aesthetics but also risk the loss of critical information, particularly in forensic analysis, medical diagnostics, and historical documentation. This study aims to compare the effectiveness of three image restoration methods—Convolutional Neural Network (CNN), Non-Local Means (NLM), and Lucy-Richardson—through a systematic literature review approach.

Design/methodology/approach: This research adopts a Systematic Literature Review (SLR) methodology, analyzing peer-reviewed articles from IEEE Xplore and other reputable sources. Each method is evaluated based on key restoration criteria, including detail preservation, noise handling, and computational complexity.

Findings/result: CNNs demonstrate superior performance in restoring semantic and complex structural details due to their deep learning capabilities, although they require large datasets and longer training times. NLM is effective in reducing noise and preserving texture details but is computationally intensive. The Lucy-Richardson algorithm, as a classical deconvolution method, is relatively simple and does not require training data, yet it heavily depends on accurate point spread function (PSF) estimation and is susceptible to noise amplification. The analysis indicates that a hybrid approach combining these methods can significantly enhance image restoration quality.

Originality/value/state of the art: This study offers a comprehensive comparative analysis of three widely used restoration techniques and highlights the potential of hybrid systems. By integrating the strengths of CNN, NLM, and Lucy-Richardson, a more adaptive and optimal restoration strategy can be developed to address diverse types of image degradation.