SA-FBCNN：一種具空間自適應性與彈性的盲式卷積神經網路於JPEG影像壓縮雜訊去除之研究

Abstract

本論文提出一種改良型JPEG影像壓縮失真修復網路SA-FBCNN（Spatially-Adaptive Flexible Blind CNN），結合FBCNN（Flexible Blind CNN）基礎架構與SAFMN（Spatially-Adaptive Feature Modulation Network）的空間自適應特徵調製機制。傳統JPEG壓縮因區塊處理方式產生明顯的方塊效應和環狀偽影，而現有深度學習方法雖有成效，但缺乏對空間特徵的自適應調整能力。本研究將FBCNN架構中的ResBlock替換為Feature Mixing Module模組，增強網路對不同尺度特徵的建模能力。接著，在對訓練完成後的模型進行分析時，我們觀察到品質因子（Quality Factor, QF）預測分支因普遍的Dead ReLU現象，其輸出對不同輸入趨於恆定。基於此，我們進行了模型剪枝，實驗證明移除該分支可在幾乎不影響效能的前提下，使參數量（60.01M）大幅減少，甚至少於原始FBCNN（71.90M）。實驗結果顯示，我們的輕量化模型在PSNR指標上平均提升約0.15dB，在圖像重建品質和細節保留方面表現更優越。This thesis proposes SA-FBCNN (Spatially-Adaptive Flexible Blind CNN), an enhanced network for JPEG compression artifact removal that integrates the architecture of FBCNN (Flexible Blind CNN) with the spatially-adaptive feature modulation mechanism of SAFMN (Spatially-Adaptive Feature Modulation Network). Traditional JPEG compression introduces noticeable blocking artifacts and ringing effects, while existing deep learning approaches often lack the ability to adaptively adjust spatial features.We enhance the network by replacing FBCNN’s ResBlocks with Feature Mixing Modules, which improve its ability to model multi-scale features. Additionally, post-training analysis revealed that the output of the Quality Factor prediction branch remains nearly constant across different inputs, likely due to the Dead ReLU phenomenon. Based on this observation, we removed the branch. Experiments confirmed that this pruning significantly reduced the model’s parameter count from 71.90M in the original FBCNN to 60.01M with negligible performance loss. The experimental results show that our lightweight model achieves an average PSNR improvement of approximately 0.15 dB over the original FBCNN, offering better image reconstruction quality and detail preservation.