The Study of Mean-Based Weights for Non-Local Means Image Denoising Algorithm

• Main Authors: Ying-Ze Chen, 陳瑩澤
• Other Authors: Jan-Ray Liao
• Format: Others
• Language: zh-TW
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/38928984848328893946

碩士 === 國立中興大學 === 電機工程學系所 === 105 === Non-local means (NLM) is one of the most effective image denoising algorithms proposed in recent years. It works by comparing the similarities between image patches. More similar patches are given higher weights and the weighted averages of image pixels can achieve excellent denoising results. However, the similarities between patches will reduce under heavy noises. In turn, this will reduce the effectiveness of denoising and cause image blurring. To alleviate the reduction of patch similarity under heavy noises, this thesis proposes a method of increasing similarity under heavy noises by using patch means to reduce noise energy through low-pass filtering. We will show that using patch means instead of the interfered patch itself in weight calculation can improve similarity calculation. In the experiments, we will analyze the effects of the following parameters on peak signal-to-noise ratio (PSNR): radius of mean calculation, weight, radius of search area, and patch size. With the additional consideration of the execution time, the best parameters by trading off between PSNR and execution time are found. We will show that our proposed method can obtain better PSNR and show less blurring when compared against previous methods.