在此研究中，主題在於數位影像的資訊藏匿技術之研究。此篇論文包含了三個主要部分。第一部分中，我們研究基於量化索引值調變(Quantization index modulation)之資訊藏匿技術。期間發展出一個新型巢狀結構的量化索引值調變之可逆式資訊藏匿技術。為了要能夠在解碼端重建原圖，在嵌入浮水印過程中產生的量化誤差也一併編碼進浮水印中。除此之外，巢狀結構也能夠提供更多的浮水印容量。在我們的實驗中，跟傳統的方法比較起來，雙層巢狀結構可以提供兩倍的資訊藏匿量。接著我們考慮發展更高透明度的演算法，而開發出一種基於離散分數餘弦變換(Fractional discrete cosine transform)的演算法。此項技術將原圖和藏有浮水印之原圖融合為一；在解碼端也可以用逆向演算法完全分離兩張影像。在我們的實驗中，跟傳統的方法比較起來，透明度提升了21.34%。第二部分中，我們的目標在於基於像素值差(Pixel difference)和直方圖修正(Histogram modification)之資訊藏匿技術。由於鄰近哈爾小波(Haar wavetlet)係數的相似性，大部分鄰近係數的差值都接近於零。根據這種特性而對現今一種演算法提出改進的方式。在我們的實驗中跟先前的方法比較起來，透明度提升了27.78%，並且多藏匿了48.48% 位元的資訊。在第三部分中，恰可察覺失真度(Just noticeable distortion)模型之運算速度是研究的主題。我們研究了傳統的恰可察覺失真模型，並提出了對於計算複雜度的改善，其運算速度提高了六倍。In this research, we focus on information hiding in digital images. The dissertation consists of three major parts. In the first part, we investigate information hiding algorithm based on quantization index modulation (QIM). A novel quantization index modulation based reversible information hiding algorithm with a nest-structure has been developed. In order to reconstruct the host image in the decoder, the quantization errors caused by QIM process are normalized and coded into the watermarked image. Besides, the capacity can be increased by taking advantage of the proposed nest-structure. Compares with the conventional QIM watermarking, a two-level nested QIM watermarking can hide 2 times secret bits in our experimental results. Next, we consider achieving a higher transparency algorithm. An algorithm based on the fractional discrete cosine transform is developed. This technique merges the host image and QIM watermarked image into one, and the host image can be exactly reconstructed by using the inverse algorithm. Experiments show that the transparency of the FDCT watermarked image can be increased. Compares with the conventional QIM watermarking, the PSNR of four test watermarked image is increased by 21.34%.In the second part, we aim at information hiding by pixel difference and histogram modification. Due to the similarity of neighbor Haar wavelet coefficient’s value, most difference between pairs of adjacent coefficients are close to zero. An improvement for an existed technique based on this characteristic has been made. Compares with the previous method, the PSNR of the six test watermarked images is increased by 27.78%, and 48.48% more secret bits hidden inside. In the third part, the computation complexity of just noticeable distortion model is the topic. We studied the conventional just noticeable distortion model and make an improvement in computation speed; the speed is about 6 times faster than that of the conventional JND model.
