高容量且高品質的高動態範圍影像資料隱藏演算法

Abstract

本論文提出了一個全新型的高動態範圍影像資料隱藏演算法，並具有高容量且高品質的特性。演算法使用紅－綠－藍－指數(RGBE)格式之高動態範圍影像，訊息之嵌入包含下列三大步驟：第一、預估嵌入量：我們首先計算掩護影像之明亮度分佈，並依據人類視覺對明亮度的靈敏程度之異同，來估計單一像素之容許嵌入量；第二、修正嵌入量：以相鄰兩像素為一組，依據其指數數值之異同，動態調整該像素之嵌入量，即對於指數數值相異者，提高預估之嵌入量；此外，我們也調整單一像素預估之嵌入量，至可容許之最大範圍；最後，若單一像素具有特殊的指數數值，我們調整允許其具有最大之容許嵌入量。經過此兩步驟，我們已經決定了單一像素之容許嵌入量。第三、分配色彩頻道嵌入量：考量人類視覺對色彩感知之差異度，我們先決定每個色彩頻道之嵌入權重，然後根據已決定之單一像素容許嵌入量，將秘密訊息動態嵌入在不同的色彩頻道。經過上述三個步驟，我們即可建構出具有高容量、高品質之偽裝高動態範圍影像。訊息之擷取可依訊息嵌入之相似流程獲得。我們使用視覺差異估測系統(Visual Difference Predictor)來量測掩護影像與偽裝影像之視覺差異程度。實驗結果顯示：本演算法在每個像素嵌入高達11.6~13.5位元之秘密訊息時，人眼無法辨認掩護影像與偽裝影像之差異。就我們所知，我們演算法的嵌入量是目前文獻中之最大嵌入量。我們的演算法具有五大特性：高容量的訊息嵌入量、高品質的偽裝影像、符合人類視覺靈敏度之可適應性訊息嵌入方式、高安全性、盲擷性。This paper proposes a high capacity and quality steganographic algorithm for high dynamic range images. Given an input high dynamic range image with the RGBE format, our algorithm considers the luminance distribution characterized in the cover image and the intensity difference existed in the neighboring pixels. The algorithm estimates the overall amount of message that can be embedded in a single pixel. Then, different red, green, and blue channels are assigned to certain parts of the overall message based on the color sensitivity inherent from the human vision. Finally, we present a novel embedding technique to embed the payloads at each color channel. This allows our algorithm to be able to embed a great number of secret messages, yet producing the stego high dynamic range image with high visual quality. Experimental results show that, on average, the embedding capacity can achieve the magnitude of 11.6~13.5 bits per pixel when using our proposed algorithm. To the best of our knowledge, this capacity is the highest in the literature of steganography for high dynamic range images. Our algorithm has the following characteristics; namely, (1) it is appropriate to characteristics of high dynamic range images, (2) it offers extremely high capacity, (3) it provides high visual quality inherent to human vision, (4) it ensures the secure steganography, and (5) it belongs to the blind detection scheme.