RIDNet深度學習去噪模型的提升:基於網絡結構與損失函數的調整

Abstract

自1970年代後期以來，隨著計算機視覺領域和數字影像處理的不斷發展，影像去噪技術也獲得了改善和創新。從最初基於空間域與變換域的濾波器、字典學習和統計模型的方法，到現今基於人工智慧的機器學習技術，可以發現影像去噪的方法日益多樣和精密。儘管許多去噪模型已經取得了相當不錯的成果，但仍然存在一些缺陷，比如需要手動設定參數、優化效果不佳，或者僅適用於特定類型的雜訊等。隨著卷積神經網路學習能力的增強和硬體技術的提升，基於深度學習的技術逐漸成為主要的影像去噪方法。卷積網路不僅能處理大量數據，還能進行高效的訓練和學習。然而，一般情況下的雜訊是無法得知的，因此能夠面對真實影像雜訊的盲去噪模型在當今的影像處理中尤其重要。這些模型必須具備強大的自適應能力，能夠有效地從影像中提取出雜訊的特徵並進行有效的去除，而不需要對雜訊進行先驗知識的設定。因此，在本篇論文中，對於盲去噪模型，我們將專注於擁有注意力機制和殘差學習的RIDNet，並對其EAM層數、激活函數及損失函數進行修改，並與其他現有的深度學習模型進行比較，如DnCNN和CBDNet。這些比較將幫助我們更了解模型，並為影像去噪技術進一步提供改善指引。Since the late 1970s, with the continuous development in digital image processing and computer vision, image denoising techniques have undergone significant improvements and innovations. From the initial methods based on spatial and transform domain filters, dictionary learning, and statistical models, to the present-day machine learning techniques based on artificial intelligence, the methods for image denoising have become increasingly diverse and sophisticated. Despite the considerable achievements of many denoising models, they still suffer from some drawbacks, such as the need for manual parameter tuning, poor optimization, or applicability limited to specific types of noise.With the enhanced learning capabilities of convolutional neural networks (CNNs) and advancements in hardware technology, deep learning-based techniques have gradually become the primary methods for image denoising.Convolutional networks can handle large volumes of data and perform efficient training and learning. However, noise in real-world scenarios is often unknown, making blind denoising models particularly crucial in contemporary image processing. These models must possess robust adaptive capabilities to efficiently extract noise features from images and perform effective denoising without requiring prior knowledge about the noise. Consequently, in this paper, we focus on the RIDNet, which incorporates attention mechanisms and residual learning for blind denoising. We aim to modify its enhancement attention modules (EAM) layer architecture, activation functions (ACT), and loss functions, and compare it with other existing deep learning models such as DnCNN and CBDNet. These comparisons will help us understand the weaknesses and strengths of model and provide further guidance for improving image denoising techniques.