使用多層感知神經網路移除影像高斯雜訊

Abstract

去除影像雜訊一直以來都是數位影像處理領域中最重要的議題之一。到目前為止，有很多基於影像先驗資訊的去除影像高斯雜訊演算法在近幾年已經發展的非常成熟。但是，在這一類的方法中，使用全域先驗資訊或是使用特定先驗資訊所開發出來的這兩種影像去除雜訊演算法卻沒有辦法同時擁有彼此的優點。基於觀察到這個現象之後，此篇論文的研究採用一個多層感知神經網路去有系統地結合這兩種分別使用全域先驗資訊和使用特定先驗資訊的去雜訊方法，以保留彼此的優點進而得到更好的影像去除雜訊效果。此篇論文提出的方法流程主要由五個步驟所構成。分別是估計雜訊強度、使用BM3D和EPLL作第一階段平行去除雜訊、把第一階段的去除雜訊影像分解出重疊的區塊、使用多層感知神經網路預測每個像素的值、聚合所有預測來還原除雜影像。多層感知神經網路再這裡扮演重要的角色。對於一個多層感知神經網路，影像除雜的問題被模擬成一個分類的問題，並且做了一系列的實驗去找出對於提出的模型最好的多層感知神經網路架構。實驗結果顯示，提出的方法流程不只在PSNR得到了改善，和目前最先進的兩個影像去除雜訊演算法(BM3D和EPLL)相比，也得到了更好的視覺品質。

Image denoising has always been one of the most important issues in the domain of digital image processing. So far, many image denoising algorithms based on learning image priors for removing Gaussian noise have evolved a lot and matured over recent years. However, in such prior-based denoising methods, using specific priors have advantages that do not exist in the methods using generic priors and vice versa. Inspired by this observation, a multilayer perceptron (MLP) neural network is adopted in this thesis to systematically combine these two image denoising methods, separately using generic and local priors, in order to keep both their own advantages and obtain a better performance. The denoising framework are mainly composed of 5 steps: noise level estimation, predenoising parallel by BM3D and EPLL, decomposing the results into overlapping patches, using a MLP for pixel-wise estimations, and aggregation for final denoised result. MLP plays a critical role in the proposed framework. We model the image denoising as a classification problem for the MLP and do a series of experiments to verify the best structure. Experiment results demonstrate that not only does the proposed framework gain improvements in terms of PSNR but also have better denoising visual quality compared to the two state-of-the-art denoising algorithms, BM3D and EPLL.