植基於權重模數技術之高動態範圍影像 資訊隱藏演算法與變動量預測機制

Abstract

高動態範圍影像是未來的潮流，目前資訊隱藏使用高動態範圍影像作為掩護影像的研究也日益受到重視。本論文提出基於權重模數的高動態範圍影像動態調整資訊隱藏演算法與權重模數預測機制演算法。 我們所提的第一個演算法為動態調整資訊隱藏演算法；該演算法以權重模數為訊息嵌入核心技術，並以RGBE格式的R、G、B色彩頻道為秘密訊息嵌入單位。我們的演算法以動態調整的方式來進一步降低偽裝影像變動量，解決像素溢位的問題。此外，我們的演算法也同時維持影像合法性，使得產生的偽裝影像能避免偽裝偵測攻擊，不至產生安全性之疑慮。實驗結果顯示，相較於事先調整方法，動態調整方式可以減少67%到91%的像素調整量；整張偽裝影像可減少0.8%到4.39%的變動量。色調映射結果顯示當每像素嵌入15位元秘密訊息時，偽裝影像之PSNR數值高於30 dB，具有良好之視覺品質。演算法可以抵抗RS偽裝偵測攻擊；掩護影像與偽裝影像之像素直方圖分析也顯示兩者具有高度之相關性。我們的演算法具有高藏量、高品質、高安全性之特性。 我們所提第二個演算法是植基於權重模數之高動態範圍影像秘密訊息嵌入預測機制。我們的預測機制考量秘密訊息分佈、訊息嵌入量與掩護影像特徵等三大因素。具言之，我們經數學分析，使用四個獨立矩陣相乘計算後，即可預知偽裝影像之期望變動量。此預測機制可讓使用者在嵌入秘密訊息前即可預知偽裝影像之嵌入結果。此外，對特定掩護影像而言，使用者如欲產生較低變動量之偽裝影像，該機制可在秘密訊息出現機率範圍內，預先求出最佳之數值。我們以低、中、高鍵值的高動態範圍影像進行實驗。實驗結果顯示：我們的預測機制具有高度之準確性，其預測誤差範圍在0.01%～0.59%之間。當實際在每像素嵌入高達15 位元之秘密訊息時，預測之偽裝影像期望變動量符合實際嵌入測得之變動量；此外，偽裝影像經兩種色調映射演算法處理，產生之低動態範圍影像具有高於30 dB之PSNR數值，仍具良好之視覺品質。 本研究具有以下貢獻：動態調整之演算法能降低像素變動並且維持影像格式合法性；預測機制可以精確預測偽裝影像變動量；預測機制可以協助使用者產生符合需求之偽裝影像。所提之方法擴展資訊隱藏技術在高動態範圍影像之應用。

High dynamic range images are able to store a greater dynamic range of luminance in order to more accurately represent the range of intensity levels found in real scenes. This thesis presents a dynamic approach for data hiding, and introduces a prediction scheme based on the weighted modulus embedding technique. We provide a dynamic pixel adjustment approach for data hiding using the weighted modulus embedding technique. In particular, we consider the R, G, and B channel of the Radiance RGBE encoding format as an embedding unit. In contrast to a static approach, our algorithm alters pixels dynamically when they are encountered with an overflow or underflow problem. This allows our method to reduce the pixel variation due to message embedding and produce a stego image that has less distortion than the static approach. In addition, our dynamic approach maintains the integrity of the RGBE image encoding. Consequently, we are able to produce a stego image which causes no suspicion when the legality of the image encoding is inspected. Experimental results show that the dynamic adjustment can reduce 61%~91% of the number of pixels and decrease 0.8%~4.39% the image variation in comparison with the static approach. Tone-mapped images perform with a good visual quality where the PSNR values are over 30 dB when each pixel is conveyed with 15 bits of secret message. Our data hiding algorithm can resist the RS steganalysis attack and provides high correlation coefficients between the pixel histograms of cover and stego images. Our algorithm provides benefits of high embedding capacity, high image quality, and high security. The second algorithm we present is a prediction scheme that is able to foresee the expected mean squared error. Our scheme considers three factors including the probability appearance of the secret bit “0” and “1,” the embedding capacity, and the medium features of the high dynamic range images. Specifically, we present a mathematical analysis and we compute the expected mean squared error by simply multiplying four independent matrices. Given a cover high dynamic range image and the probability of the secret bits, our scheme can forecast the mean squared error prior to the real message embedding. Given a range of probability appearance, our mechanism can suggest the best values with this range for a specific cover image in order to produce the stego image that has the smallest pixel variation. We include a variety of high dynamic range images including low, middle, and high keys when conducting an experiment. The experimental results show that our scheme reveals a high accuracy of prediction, the error rates being in the range of 0.01%~0.59%. The accuracy is preserved even when each pixel is concealed with 15 bits of secret messages. Tone-mapped images produced by two different tone mapping algorithms demonstrate that the PSNR values are over 30 dB, and produce a good visual quality of stego images. In conclusion, this study provides three contributions: the dynamic adjustment approach which reduces the pixel variation and maintains the image integrity; the prediction method which performs with accuracy; the prediction scheme which helps users generate the stego image satisfying desirable demands. Our schemes expands applications of data hiding for high dynamic range images.