RST不變性數位浮水印技術

Abstract

網際網路的發達，使得聲音、影像及影片等多媒體資料容易被取得，甚至於遭到有意及無意的破壞與盜用。為了使多媒體資料原創者著作權受到保護，不論影像遭到旋轉、縮放及位移處理後，仍可以準確辨識該影像著作權，本研究以適用於浮水印之尺度(Scale)不變特徵值轉換，擷取影像中較穩定的特徵值，並分別在影像空間域及頻率域上，配合浮水印量化及中頻係數選取，在特徵值周圍進行嵌入與萃取浮水印訊息，以有效抵抗RST攻擊。 實驗結果顯示，影像經過旋轉(Rotation)攻擊，例如旋轉30度以上，或是受到放大2倍以上之縮放(Scaling)攻擊，其整體浮水印萃取效能(Performance)為75%及89%。從結果可知，不論在空間域或是頻率域浮水印技術，我們都可使浮水印萃取之效能較為強韌。

Traditional watermarking schemes are sensitive to geometric distortions, in which synchronization for recovering the embedded information is challenging due to the disorder caused from rotation, scaling or translation (RST). This thesis presents robust watermarking algorithms based on SIFT features, which are resilient to geometric attacks, including RST, and even warping. By embedding watermark bits in RST-invariant keypoints in the image and aligning the embedding direction according to the principle axis, watermark can be well preserved and extracted even if the image is destroyed by RST. Since limited robust features can be found in the spatial domain, deriving a robust spatial-domain watermarking is another difficult problem. However, besides the frequency-domain watermarking schemes, we also design a quantization watermarking scheme in the spatial domain, which is robust to RST. Experiment results show the proposed algorithms are superior to many existing works.