以量化為基礎之數位影像浮水印

Abstract

網路與高容量儲存設備使得數位資訊內容(如：音樂、影像與視訊…等)的交換與傳播更加便利，但卻也讓著作權的保障面臨嚴厲的挑戰，進而影響創作的發展與進步。另外，由於軟體科技日新月異，幾可亂真的竄改使得判別數位資訊內容的真偽更為不易。浮水印(Watermarking)技術被廣泛地討論與研究，即是寄望該技術的成熟，能為上述種種挑戰提供解決之道，並為數位權利的管理(Digital Rights Management)提供堅實的基礎。 本論文首先探討數位影像浮水印技術的基本架構及其各組成要件的功能。在評估浮水印技術時，透明度(Transparency)代表浮水印資訊的嵌入對視覺品質的影響程度；強健性(Robustness)是指所嵌入的資訊抵抗惡意攻擊或各種信號處理的能力；容量(Capacity)則為載送數位浮水印資訊量的能力。以上評量參數在不同的浮水印應用時，有不同程度的要求。本研究列舉各種浮水印應用，並對各項參數的要求提出建議。 早期對數位浮水印技術的探討，常將欲嵌入的資訊視為數位通訊系統的傳送訊息。此訊息在到達接收端前，會經歷原始影像的干擾與傳輸通道的攻擊。常被提及的展頻浮水印(Spread Spectrum Watermarking)技術在最初被提出時，即以此觀點進行分析與討論。然而，此觀點忽視了原始影像在傳送端為己知的事實，若能在嵌入數位浮水印時參考原始影像，來自原始影像的干擾應能因而減少。以量化為基礎的浮水印技術(Quantization-Based Watermarking)即以此概念來設計。此技術之特點為在接收端毋需原始影像即可萃取所嵌入的浮水印資訊(Blind Detection)，該技術可以量化索引值調變(Quantization Index Modulation, QIM)之方法為代表。本論文除了分析與討論以純量量化(Scalar Quantization)為基礎的數位影像浮水印外，並討論與比較各種改善機制。實驗結果顯示，各種改善機制雖有相當的效果，但應用的場合均有所限制。本研究因而提出非中心量化(Non-Centric Quantization)的方法，以此法嵌入浮水印的效能表現，在大部份的應用場合中均優於傳統的量化浮水印技術。 本論文亦討論以向量量化(Vector Quantization, VQ)為基礎的數位浮水印技術。向量量化技術以其簡單的解碼器(Decoder)見長，若在編碼時將浮水印資訊納入考量，則可同時達成浮水印與壓縮的功能(Joint Watermarking and Compression, JWC)。論文中採用二元樹分割法(Binary Tree Splitting Method)產生向量量化中所需的碼簿(Code Book)，因而在編碼的同時即可嵌入浮水印。實驗結果顯示此方法在各方面的效能均優於其他以向量量化為基礎的浮水印技術。而其簡單快速的運算量，更大幅提昇了可應用性。 現今的影像浮水印技術，大部份都針對灰階的原始影像處理。提及彩色影像者，亦建議以灰階影像的浮水印技術個別地應用於彩色影像中的三通道(Tri-Channel)信號或其所屬的灰階信號。此概念忽略了三通道信號彼此的關聯性。本研究發現，將浮水印資訊嵌入各通道後的效能表現，將會因其遭遇攻擊的種類不同而有所差異，使用者將因此無法決定應將浮水印嵌入至哪一通道。費伯納契晶格(Fibonacci Lattice)曾被提出應用於色彩量化(Color Quantization)的研究。除了其優異的視覺效果外，解碼端只需利用少量參數，即可簡單地產生與編碼端相同的調色盤(Color Palette)。本論文提出費伯納契晶格索引值調變(Fibonacci Lattice Index Modulation, FLIM)的方法。以此法將具有驗證功能的易碎(Fragile)浮水印嵌入到原始影像的彩色信號成分，將有相當優異的透明度表現。若再配合QIM技術，把具有保護所有權功能的強健(Robustness)浮水印嵌入到原始彩色影像的灰階信號成分，將可使數位影像同時載有二種不同功能的浮水印資訊。本研究實驗結果顯示：易碎浮水印的嵌入並無損於強健浮水印的效能。此一混合浮水印技術，將大幅提高浮水印技術的可應用性。

The computer networks and mass storage devices make the exchange and spread of digital multimedia content more easy than before, yet challenge the copyright protection and obstruct the development of creative work. In addition, many fantastic software can tamper the digital content imperceptibly, and thus raise the difficulty of authentication. Consequently, watermarking techniques are extensively discussed with the aim to address these issues and provide the DRM (Digital Rights Management) with a firm basis. This dissertation presents a preliminary discussion of watermarking systems. Each part of a generic watermarking scheme is discussed essentially. In the invisible watermarking systems, three important properties: transparency, robustness, and capacity, are typically evaluated to criticize the performance of a watermarking system. However, these properties are demanded different extents depending on what applications the watermarking systems are applied. Many researches regarded the watermark as the message to be sent over a digital communications system, and the transmitting will inevitably sustain the interference from the host signal and the superimposed noise from attacks. However, this viewpoint ignores the fact that the host signal is known to the watermark embedder. The quantization-based watermarking has the inherent advantage that the host signal is taken into account for rejecting the interference from it, where the QIM (Quantization Index Modulation) is the representative approach. This study analyzes and compares a variety of improving skills for quantization-based watermarking. The experimental results have shown that those improving skills more or less accomplished their expectation in some restricted scenarios, but not for all the cases. In this dissertation, the non-centric quantization is proposed to improve the robustness of quantization-based watermarking. The simulation results have demonstrated that the proposed scheme outperforms the convectional quantization-based watermarking for most of the applicable cases. This study also presents the watermarking schemes based on vector quantization (VQ), which inherently have the feature of joint watermarking and compression (JWC). A modified binary tree splitting method (BTSM) is proposed to generate the codebook for VQ, and the watermark message is thus embedded by the replacement of the representative index when the host vector is encoded. The performance superiority over other watermarking schemes based on VQ is presented. Notably, the lightness of computation cost of the proposed scheme enhances its applicability. Most of the watermarking approaches are proposed for the host images in gray-level. As for the color host images, the most intuitive approach is directly applied the methods designed for gray-level images to the tri-channel signals of the host images. This study evaluates the performances of the watermark messages embedded into different channels of the color host image. The experimental results have shown that the answer to 'which channel of the color host image is suitable for embedding watermark message?' is indecisive, yet should be answered according to the attack to which the watermarked images will resist. The color quantization based on the Fibonacci lattice has been proven to be competitive to other approaches for its good visual quality and the simplicity of color palette generation at decoder end. A novel image watermarking approach, Fibonacci Lattice Index Modulation (FLIM), is specifically proposed for the color image watermarking in this study. FLIM unobtrusively embeds a fragile watermark into the chromatic component for the purpose of authentication. With another robust watermark embedded into the luminance component by the QIM approach, the hybrid watermarking system embeds two watermark messages with different purposes concurrently, and can thus be applied to a variety of scenarios.