基於可逆神經網路與正交轉換之可變率影像壓縮方法;Variable Rate Image Compression Based on Invertible Neural Networks and Orthogonal Transforms

NCU Institutional Repository > 資訊電機學院 > 資訊工程研究所 > 博碩士論文 > Item 987654321/98572

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: https://ir.lib.ncu.edu.tw/handle/987654321/98572

题名:	基於可逆神經網路與正交轉換之可變率影像壓縮方法;Variable Rate Image Compression Based on Invertible Neural Networks and Orthogonal Transforms
作者:	呂珮伶;Lu, Pei-Ling
贡献者:	資訊工程學系
关键词:	影像壓縮;可逆神經網路;正交轉換;深度學習;Image Compression;Invertible Neural Network;Orthogonal Transform;Deep Learning
日期:	2025-08-14
上传时间:	2025-10-17 12:56:36 (UTC+8)
出版者:	國立中央大學
摘要:	可逆神經網路（Invertible Neural Networks, INNs）的可逆特性能避免處理前後的潛在資訊損失，基於INN的影像壓縮方法因此被提出以在高壓縮率下提升重建影像品質。然而，現有方法所採用的像素混合下採樣方式缺乏特徵去相關能力，且相關模型的壓縮位元率調整並無彈性，需要針對不同的率失真（Rate-Distortion）表現訓練個別模型，本研究因此提出結合正交轉換與單一模型可變位元率控制機制的可逆神經網路架構。首先，我們將原本的像素混合下採樣層替換為具有正交特性的轉換，在提升特徵去相關性的同時亦保留INN的可逆性，實現更有效率的頻域多尺度表徵能力。其次，本研究整合近期所提出的可變編碼率方法，透過單一參數動態調整壓縮位元率，提升模型的部署彈性。我們亦改進相關方法的耦合層（Coupling layer）架構，透過引入殘差密集塊（Residual Dense Block），有效捕捉頻率域與空間域間的依賴性，進一步提升壓縮效能。實驗結果顯示，本研究所提出的架構在影像資料集Kodak與CLIC-professional 上，相較現有基於INN的影像壓縮模型取得明顯的率失真性能提升，並進一步測試以驗證各種正交轉換方式的效果，顯示目前方法的普遍適用性與未來改進的潛力。;The invertibility of Invertible Neural Networks (INNs) effectively prevents potential information loss during processing. Image compression methods based on INNs have thus been proposed to enhance reconstructed image quality at high compression ratios. However, existing approaches predominantly utilize pixel shuffling for spatial downsampling, which lacks the ability to decorrelate features effectively. Additionally, these methods lack flexibility in adjusting compression bitrate, requiring separate model training for different rate-distortion (R-D) performances. To address these limitations, this study proposes an INN-based image compression framework that combines orthogonal transforms with a single-model variable-rate control mechanism. First, we replace the conventional pixel shuffle downsampling layer with orthogonal transforms, enhancing feature decorrelation while preserving the inherent invertibility of the INN architecture, thereby achieving a more efficient multi-scale frequency representation. Second, this research integrates a recently proposed variable bitrate method, dynamically adjusting the compression rate via a single parameter, significantly improving model deployment flexibility. Furthermore, we refine the coupling layer architecture by incorporating Residual Dense Blocks (RDB), effectively capturing dependencies between frequency and spatial domains and further improving compression performance. Experimental results demonstrate that the proposed framework achieves significant improvements in rate-distortion performance over existing INN-based methods on the Kodak and CLIC-professional image datasets. Additional experiments validate the effectiveness of various orthogonal transforms, highlighting the broad applicability and potential for future advancements of the proposed approach.
显示于类别:	[資訊工程研究所] 博碩士論文

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	5	检视/开启

在NCUIR中所有的数据项都受到原著作权保护.

社群 sharing

数据加载中.....