具有低失真性之黑白影像藏密學技術

Abstract

本論文提出一種新型且能夠有效的減少因隱藏秘密訊息所造成視覺雜訊的黑白影像資訊隱藏技術，我們所提出的方法包含了三個部分，首先是一張以0或1像素值所構成的黑白掩護影像，將它切割成子影像區塊，在每一個子影像區塊中，我們將設計一種簡單又有效率的像素可變動性機制，即檢查中心點像素的每一個鄰居，以評估中心點像數的可變動分數，若中心點像素與它附近鄰居像素的黑點或白點有較多的差別時，我們則給予此像素較高的可變動性分數。第二部分，我們利用Arnold轉換將黑白影像擾亂成亂碼影像，再去檢查每一個子影像區塊中，是否存在一個以上的可變動像素，如果不存在，則必須再以Arnold轉換擾亂，直到每一個子影像區塊存在一個以上的可變動像素為止。最後，我們提出一種創新的訊息嵌入技術，每一次取3個子影像區塊做為資訊隱藏的單位，每一個單位將一次嵌入3個位元的秘密訊息，並且，在訊息嵌入的過程中，最多只有2個子影像區塊的可變動像素會被修改。本論文所提出的方法是不需要原始掩護影像就可以將秘密訊息取出，諸多的實驗結果也証明新方法所產生在偽裝影像上的雜訊，以人類眼睛無法察覺到異樣。最後，與現存的方法比較，新方法所得出的偽裝影像失真性最小，平均修改機率僅為0.333位元／像素。In this paper, we propose a novel data hiding scheme for binary image which is capable of significantly reducing the visual artifacts caused by secret data embedding. Our new scheme includes three parts. Firstly, the cover image represented by a binary sequence of 0 or 1 pixel values is subdivided into blocks of sub-images. For each block we introduce a simple yet efficient scheme to estimate a pixel's ”flippable value” by examining its neighbor. We give high ”flippable value” to a pixel where its neighbors are much different from the current pixel value being examined. Secondly, we utilize the Arnold transform to shuffle the binary image into a noise image. The transformation is conducted several times until each block of sub-image contains at least one pair of pixels with ”flippable value,” which indicates that this block of sub-image is feasible to embed a bit of secret message by flipping the corresponding pixel value. Finally, we present a novel data embedding technique, which employs three blocks of sub-image as a basic embedding unit. This technique ensures the basic unit to convey 3 bits of secret message by at most flipping two pixels with ”flippable values” on any of two blocks. Our scheme belongs to the blind extraction where the secret data can be extracted without referring to the cover image. Experimental results show the stego image produced by our scheme is imperceptible to the human eyes. Comparing to existing schemes, our method produces the least image distortion, being 0.333 bits per pixel.