使用倒傳遞類神經網路優化於空間域與頻率域之數位浮水印辨識

Using Back-Propagation Neural Network Optimization for Digital Watermarking Identification of the Spatial Domain and Frequency Domain

黃義傑(Yi-Jie Huang)

指導教授 ： 顏志達

國立虎尾科技大學/電機資訊學院/電機工程研究所/碩士(2014年)

https://doi.org/10.6827/NFU.2014.00280

摘要

數位浮水印加密技術是一種常用的方法來保護智慧財產權和版權。儘管浮水印是一個強健性的方法來保護產權，但是環境干擾圖像或者是透過網路傳播的人為修改也可以使水印受到破壞。為了能夠更強健的保護水印以及取出水印，於研究論文中，將探討空間域與頻率域嵌入的水印遭受破壞時，採用倒逆傳遞類神經網路演算法來還原嵌入的浮水印。
本篇論文主要是針對已被雜訊干擾的浮水印做辨識，我們使用32 x 32的校徽圖章作為浮水印辨識碼，分別使用三種方法將浮水印嵌入256 x 256的Lena主圖中，第一種方法是空間域中最常使用的最低位元替代法，另外兩種是頻率域中的離散餘弦轉換以及離散小波轉換，比較空間域與頻率域嵌入浮水印後遭受到自然破壞或者是影像幾何破壞的峰值訊號雜訊比(PSNR)以及提取出浮水印的正規化相似性(NC)。模擬自然破壞的雜訊我們使用胡椒鹽雜訊與高斯雜訊；模擬影像幾何破壞則是中值濾波、剪裁、旋轉、JPEG壓縮。包含浮水印的三張不同嵌入法的Lena主圖分別遭受以上的雜訊一定程度的破壞，初步提取出的浮水印勉強能看出相似於原始浮水印，但破壞程度過高時則無法辨識，為了提升浮水印的正規化相似性，我們使用倒傳遞類神經網路做最佳化辨識，發現經過類神經辨識出的浮水印能更相似於原始嵌入的浮水印。
空間域的最低位元替代法遇到隨機分布的雜訊則有較好的抵抗能力，例如胡椒鹽雜訊，但是遇到全面性的破壞則無任何強健性可言，雖然經過類神經的優化，能濾除一些雜訊使得辨識度提升，前提是只要破壞的像素點不影響到嵌入的像素位置則能夠順利的取出；而另外兩種頻率域浮水印嵌入法則是遇到胡椒鹽雜訊以及剪裁破壞較無抵抗能力，但是經過提出的類神經優化辨識之後水印辨識度提升幅度頗高。不管是空間域或頻率域的嵌入法，遭受嚴重的剪裁破壞時，此時已經把藏匿水印的部分剪去，就算經過優化辨識演算法，其辨識度只提升了些許程度。

關鍵字

數位浮水印 ； 離散餘弦轉換 ； 離散小波轉換 ； 胡椒鹽雜訊 ； 高斯雜 ； 倒傳遞類神經網路

並列摘要

The digital watermark is a popular way to protect intellectual properties and copyrights. Although the digital watermark is robust, environmental interference to images or artificial modification through network can ruin it. As the damage exceed the limit, the digital watermark can’t be extracted.

This thesis talks about the identification of noise-interfered digital watermarks. We use a 32x32 image of our school’s badge as the watermark identification code. Three methods are used to embed this watermark into a 256x256 image of Lena. The first method is the least significant bit(LSB) replacement, which is the most popular method is spatial domain. The other two are discrete cosine transform(DCT) and discrete wavelet transform(DWT) in frequency domain. We compare the peak signal to noise ratio(PSNR) of the naturally damaged image and the geometrically distorted image whose watermark is embedded in spatial domain or frequency domain. And we extract the normalized correlations(NC) of the watermarks.

To emulate the noise of natural damage, we introduce the Pepper and Salt noise and the Gaussian noise; To emulate the noise of geometric distortion, we introduce the median filter, clipping, rotation, and the JPEG compression. These three image of Lena with embedded watermarks are damaged to a certain level. The extracted watermarks can be barely recognized as the original watermark. But the overly damaged noises can’t be recognized. To enhance the normalized correlation(NC) of the watermark, we use the back-propagation neural network to optimize the recognition process. And we find that the watermark, which is recognized by the back-propagation neural network, is more identical to the original one.

並列關鍵字

Digital watermarking ； Discrete Cosine Transform(DCT) ； Discrete Wavelet Transform(DWT) ； Salt Pepper noise ； Gaussian noise ； Median filtering ； Back-Propagation Neural Network

參考文獻

[1]. Stephane Bounkong, Boremi Toch, David Saad, David Lowe, 2003 “ICA for Watermarking Digital Images,” Journal of Machine Learning Research 4, pp:1471-1498

[4]. I. Pitas, 1996 “A method for watermark casting on digital images,” In Proceedings of IEEE Int. Conf.on Image proceession, vol.3, pp.215-2188, Sept

[6]. G. Voyatzis and I. Pitas , 1996. “Applications of toral automorphisms in image watermark,” in Proceedings of the IEEE Int.Conf. on Image proceeding, vol.2, pp.2237-240, Sept

[7]. J. F Delaigle and C. De Vlesschchouwer, B. Macq 1997 ,“Low cost perceptive digital picture watermarking method,” in Proc.EECMSDY 97,Milan,Italy,pp 153-1167, May

[9]. M. Kutter, F. Jordan, and F. Bossen, 1997 “Digital signsture of color images using amplitude modulation” in Proc.SPIE-EI97, pp.518-526