Blind colour image watermarking techniques in hybrid domain using least significant bit and slantlet transform

Abstract

Colour image watermarking has attracted a lot of interests since the last decade in tandem with the rapid growth of internet and its applications. This is due to increased awareness especially amongst netizens to protect digital assets from fraudulent activities. Many research efforts focused on improving the imperceptibility or robustness of both semi-blind and non-blind watermarking in spatial or transform domain. The results so far have been encouraging. Nonetheless, the requirements of the watermarking applications are varied in terms of imperceptibility, robustness and capacity. Ironically, limited studies concern on the authenticity and blind watermarking. Hence, this study presents two new blind RGB image watermarking techniques called Model1 and Model2 in hybrid domain using Least Significant Bit (LSB) insertion and Slantlet Transform (SLT). The models share similar pre-processing and LSB insertion stages but differ in SLT approach. In addition, two interrelated watermarks known as main watermark (MW) and sub-watermark (SW) are also utilized. Firstly, the RGB cover image is converted into YCbCr colour space and then split up into three components namely, Y, Cb and Cr. Secondly, the Cb component is selected as a cover for the MW embedding using the LSB substitution to attain a Cb-watermarked image (CbW). Thirdly, the Cr component is chosen and converted into the transform domain using SLT, and is subsequently decomposed into two paths: three-level sub-bands for Model1 and two-level sub-bands for Model2. For each model, the sub-bands are then used as a cover for sub-watermark embedding to generate a Cr-watermarked image (CrW). Following that, the Y component, CbW and CrW are combined to obtain a YCbCr-watermarked image. Finally, the image is reverted to RGB colour space to attain the actual watermarked image (WI). Upon embedding, the MW and SW are extracted from WI. The extraction process is similar to the above embedding except it is accomplished in a reverse order. Experimental results which utilized the standard dataset with fifteen well-known attacks revealed that, among others: Model1 has produced high imperceptibility, moderate robustness and good capacity, with Peak Signal-to-Noise Ratio (PSNR) rose to 65dB, Normalized Cross Correlation (NCC) moderated at 0.80, and capacity was 15%. Meanwhile, Model2, as per designed, performed positively in all aspects, with NCC strengthened to 1.00, capacity jumped to 25% and PSNR softened at 55dB but still on the high side. Interestingly, in terms of authenticity, Model2 performed impressively albeit the extracted MW has been completely altered. Overall, the models have successfully fulfilled all the research objectives and also markedly outperformed benchmark watermarking techniques.