基於影像內插與權重模數之資料隱藏演算法之研究

Abstract

本論文針對灰階影像，提出了兩個創新的資料隱藏演算法。文獻審閱時我們發現Jung等人之演算法在特定情況下無法正確取出秘密訊息，因此我們先修正其演算法使之能正確無誤的取出秘密訊息。接著，我們提出一個創新的資料隱藏演算法。這個演算法是基於MED影像內插並混合使用OPAP與LSB-Matching Revisited (LSB-MR)之資訊嵌入方法。藉由理論分析每個像素的期望變動量，證實我們的演算法可以在固定偽裝影像品質下，預測該影像所能容許的最高藏量；我們的演算法也可以在固定嵌入量下，預測該影像所能產出之最高偽裝影像品質。實驗結果顯示：使用我們提出的演算法來做影像內插，其平均影像品質約可達29.25 dB，較文獻上Jung等人的方法平均高出4.86 dB。此外，我們的演算法比Jung等人之方法能提供更高之嵌入量，其提高之數值介於4.86-118.10%間。我們的演算法所產生之偽裝影像品質也優於Jung等人之方法，其數值介於1.39-13.76dB間。最後，我們的方法可有效抵抗RS偽裝偵測攻擊。 我們針對Li等人之資料隱藏演算法(WM-M2)之缺失，提出本論文的第二個資料隱藏演算法。我們的演算法包含三個具體貢獻；首先，我們求出額外的權重，不僅接近最佳權重也能達到更廣的嵌入量範圍，使得WM-M2演算法能夠一般化。此更彰顯以權重為基礎的資料隱藏演算法之實值貢獻。實驗結果顯示：我們求出的額外權重與理論的最佳的權重所產生的偽裝影像，兩者之PSNR平均差異僅為0.1 dB，此代表我們所求出之權重已非常接近於理論之最佳權重。其次，我們提出WM-MA，此方法為模全部狀態數，是權重嵌入技巧之變形。其狀態數是由單一像素最大變動量所產生。我們也找出用於WM-MA之有效權重，可達到平均每個像素嵌入量範圍為1.585 - 4.858個位元。此外，我們理論分析WM-M2、WM-MA以及HYWM，其中HYWM為混合使用WM-M2、WM-MA的方法。藉由理論分析可以知道在做資料隱藏的時候使用何種演算法能夠達到三者中最佳效果。除此之外，我們估算WM-M2演算法在滿足使用者需求之PSNR或是嵌入量時，其產生偽裝影像之嵌入量或是視覺品質之下限。最後，我們使用WM-M2演算法進一步改進我們的第一個資料隱藏演算法。實驗結果顯示：嵌入量比我們第一個提出的方法增加62.23 - 81.24%；偽裝影像品質比我們第一個提出的方法增加0.82 - 10.41 dB。 在本論文中，我們修正Jung等人方法無法完全正確取出秘密訊息之缺失。我們針對灰階影像提出兩個資料隱藏演算法，而且實驗結果證明我們方法之可行性。我們認為本論文提出之演算法對於應用在影像之資料隱藏領域有實質具體之貢獻。

This thesis proposes two novel data hiding algorithms for gray scale images. The first algorithm involves the correction of a faulty algorithm proposed by Jung et al. In particular, we examine the MED image interpolation technique utilized to enlarge a small scale of the cover image. We then employ a hybrid data embedding approach taking advantage of the optimal pixel adjust process technique (OPAP) and the least significant bit matching revisited (LSB-MR) technique. We present an expected variation prediction (EVP) technique which allows us to analyze the expected variation per pixel caused by the hidden message. Given a desired visual quality of the stego image, the EVP technique can predict the maximal capacity prior to the real data embedding. Alternatively, the EVP technique can foresee the optimal visual quality of the stego image which satisfies the desired embedding capacity requested by a user. Experimental results verify that the MED image interpolation method provides a good visual quality when producing a large-scale of cover image. The average PSNR value is 29.25 dB which is 4.86 dB higher than that produced by Jung et al.'s method. Our hybrid data hiding algorithm can provide a higher embedding capacity where the increased range is between 4.86% and 118.10%. Furthermore, our method provides higher visual quality where the increased range is between 1.39 dB and 13.76 dB. Finally, our method can withstand the RS steganalysis attack. We propose the second data hiding algorithm which is motivated by the insufficiency of Li's weighted modulus (WM-M2) scheme. Our algorithm is a generalization of the WM-M2 scheme. In particular, we produce effective weights which are not only close to the optimal weights but also provide a variety of embedding capacities. Experimental results confirm that these effective weights can produce corresponding stego images which are 0.1 dB less than those generated by the optimal weights. We present a diverse scheme of weighted modulus, WM-MA, which operates the modulus computation with respect to all of the available statuses. They are generated by the maximal pixel variation granted at each pixel. We also determine effective weights used for WM-MA, achieving the embedding capacities between 1.585~ 4.858 bits per pixel. In addition, we present a theoretical analysis for schemes of WM-M2, WM-MA, and HYWM, a hybrid method which uses MW-M2 and WM-MA. This allows us to suggest which scheme shall be employed in order to achieve the best performance for data hiding. Besides, we derive the lower bound of the embedding capacity/visual quality that can be produced for a stego image when employing the WM-M2 scheme to satisfy the desired PSNR/capacity requested by a user. Finally, we refine our first data hiding algorithm using the WM-M2 scheme. Experimental results show the capacity increases to the magnitude of 62.23~81.24%, and the increase of the visual quality is between 0.82 and 10.41 dB. In this study, we rectify the faulty algorithm of Jung et al. in order to ensure the correctness of the secret message extraction. We propose two data hiding algorithms for gray scale images, and experimental results demonstrate the feasibility of our schemes. We believe the significant performance provided by our algorithms offers a substantial contribution to image data hiding applications.