植基於人類視覺系統、類神經網路及支向量機之強靭影像浮水印

Robust Image Watermarking Using Human Visual System, Neural Networks and Support Vector Machines

劉騏鋕(Chi-Chih Liu)

指導教授 ： 蔡鴻旭

國立虎尾科技大學/管理學院/資訊管理研究所/碩士(2006年)

https://doi.org/10.6827/NFU.2006.00003

摘要

本論文提出三個浮水印技術來保護數位影像內容，分別為IWNN (Image watermarking based on neural networks)、DHIW (Decision-based hybrid image watermarking)及SVMLIW (SVM-based lossless image watermarking)。三者技術分別都使用到離散小波轉換(discrete wavelet transform)。在IWNN技術，利用HVS(human visible system)中JND(just noticeable difference)控制浮水印藏入的強度，並使用類神經網路(artificial neural network)中MLP(multi-layer perceptions)來記憶原始係數值及藏入浮水印後的係數值之間的關係。因此，藏匿完的浮水印影像具有高透明性並且抽取浮水印時不需使用到原始影像。由於在IWNN技術中處理影像紋理較複雜(texture)時，會影響類神經網路一般化(generalization)能力，故本論文提出DHIW技術，藉由兩個係數值關係來藏入浮水印及結合IWNN優點，以提昇整體浮水印抽出的準確率。
SVMLIW技術不修改數位原始影像下，可同時保護數位原始影像及植入擁有者(owner)簽章。主要找出小波係數中一些不變量的特徵值造出由原始影像所產生的浮水印。接著，以互斥或(XOR)方式來承載使用者的簽章以產生載體資訊(carried information)。再者，支向量機(support vector machine)用來記憶由原始影像所產生的浮水印及載體資訊兩者關係。最後，使用受訓練後的支向量機還原出使用者的簽章，以驗証原始影像。在模擬實驗結果下，本論文所提出IWNN技術、DHIW技術及SVMLIW技術能有效抵抗常見的影像攻擊，比起以往所提出的浮水印技術有更佳的效果。因此，這些技術能被有效運用至數位多媒體著作權保護及所有權鑑定。

關鍵字

類神經網路 ； 支向量機 ； 人類視覺系統 ； 正好查覺差異 ； 離散小波轉換 ； 影像驗証 ； 資料隱藏 ； 浮水印

並列摘要

The thesis presents three image-watermarking techniques, the IWNN (Image watermarking based on neural networks), the DHIW (Decision-based hybrid image watermarking) and the SVMLIW (SVM-based lossless image watermarking), to protect image copyrights. These techniques are developed in the wavelet domain. The IWNN technique employs the just noticeable difference (JND) profile, a characteristic of the human visual system (HVS) model, to control watermark-embedding strength, and utilizes an artificial neural network (ANN) to memorize the relationships between a set of the original wavelet blocks and its watermarked version. The IWNN technique makes watermarks further imperceptible, and doesn’t need the original image during watermark extraction. A drawback of the IWNN technique is that the ANN performs poor generalization ability if an image has highly complex textures, for example, Baboon image. This inspires us to propose the DHIW technique to overcome the drawback. A scheme hides a watermark bit in two coefficients of a wavelet block with complex textures. Therefore, a decision-based concept is applied to devise the DHIW technique which combines the IWNN technique with the scheme. As a result, the DHIW technique significantly improves the performance of the IWNN technique.
These two techniques mentioned above still degrade the visual quality of watermarked images due to some modifications to images during watermark embedding. A lossless image watermarking method, called the SVMLIW technique, is proposed to retain high visual quality of protected images. Although the SVMLIW technique doesn’t modify the original image, it can hide the owner’s information in an image by using the XOR operation and a support vector machine (SVM). A design concept is to generate an image-dependent watermark, a sequence of the invariant features of the wavelet blocks of an image. The SVMLIW technique applies the XOR operation to the image-dependent watermark and an owner’s signature to generate the corresponding carried information. Subsequently, a support vector machine is utilized to memorize the relationships between the image-dependent watermark and its corresponding carried information. During watermark extraction, the SVMLIW technique exploits the trained SVM to estimate the corresponding carried information without the original image, and then retrieves the owner’s signature by applying the XOR operation to the image-dependent watermark and the estimated carried information. Numerous computer simulations demonstrate that these three schemes are definitely robust to resist the common image-processing attacks, and that are significantly superior to other existing methods. Therefore, they can be effectively applied to protect multimedia contents for intellectual property and ownership identification.

並列關鍵字

Artificial Neural Networks ； Support Vector Machine ； Human Visible System ； Just Noticeable Difference ； Discrete Wavelet Transform ； Image Authentication ； Data hiding ； Watermarking

參考文獻

[1] M.-U. Celik, G. Sharma, E. Saber and A.-M. Tekalp, "Hierarchical watermarking for secure image authentication with localization," IEEE Trans. on Image Processing, vol. 11, no. 6, pp. 585-595, 2002.

[2] A. Noore, N. Tungala and M. M. Houck, "Embedding biometric identifiers in 2D barcodes for improved security," Computers & Security, vol. 23, issue 8, pp. 679-686, 2004.

[5] F.-Y. Shih and Y.-T. Wu, "Enhancement of image watermark retrieval based on genetic algorithm," Journal of Visual Communication and Image Representation, vol. 16, pp. 115-133, 2005.

[6] Y. Wang and A. Pearmain, "Blind image data hiding based on self reference," Pattern Recognition Letters, vol. 25, no.15, pp.1681-1689, 2004.

[7] J.-M. Shieh, D.-C. Lou and H.-K. Tso, "A robust watermarking scheme for digital image using self similarity," Information Technology: Research and Education, pp. 115-119, 2005.