Colour image watermarking in real life Konstantin Krasavin University of Joensuu, Finland ABSTRACT: In this report we present our work for colour image watermarking in different domains. First we consider mobile applications of watermarking and their properties. We implement a watermark into multimedia message and extract it. Using a set of observers we study image quality changes. Interesting results were found for human vision sensitivity to image type. Then we extend watermarking techniques from conventional to spectral images. We utilize wavelet domain for watermark embedding and apply different compression ratios and various illuminators to see how they effect to quality of the image and extracted watermark.. INTRODUCTION: In this report we present our work on colour image watermarking. Digital watermarking is a technique, where an identifier signal is embedded into an information carrying signal []. The watermark is embedded in such a way, that it does not disturb the information in normal conditions. The embedded watermark can be extracted for source identification. Watermarking is based on the feature of a human vision system which has different sensitivity to different frequency bands. The sensitivity of the human visual system to specific frequencies depends on screen properties, image size, and viewing distance.. IMAGES ON MOBILE DEVICES. Images on mobile devices During last years technologies in mobile devices have been progressing very fast towards better imaging experience. When first multimedia messages appeared they had very poor visual quality of images. Due to hardware limitations at that time, mobile devices had low megapixel cameras with simple optics. Typical image resolutions were 8x8 pixels with the maximum resolution of 64x48 pixels. In additional to camera limitations, displays were small and had low resolutions as well. This led to poor imaging experience on mobile device itself. When images taken by mobile phone were viewed on a PC screen, users were not always satisfied neither. Even thought PC screens are able to reproduce high quality digital images, they could not improve the image quality of original image. In some cases the effect was just opposite. PC screens were highlighting all the problems in poor image quality and picture zooming led to decreasing image quality even more. Modern mobile devices have overcome these difficulties with mobile imaging. They offer good experience of point and shoot use cases for typical users. Often cameras have a high number of megapixel, with good quality lens and xenon flash. It is clear that quality of mobile images is approaching very fast towards digital still cameras. Mobile displays are getting better as well, though they do not catching up with megapixel race.. Watermarking in mobile devices With mobile devices we have different requirements to watermarking comparing to desktop environment. As we have discussed in previous work, the most differences are coming from the fact that captured images have small size and are displayed on small size display []. Also hardware limitations on memory size, speed and power consumption are different comparing to desktop. Image type is also important for mobile usage. We performed visual quality assessment for original and watermarked images displayed at mobile phone, PDA and desktop screen (Fig.). Visual quality of images was evaluated by a set of observers using a modified subjective mean opinion score (Fig.). 37
Fig.. Watermarked images, top row, original images, bottom raw. Score MOS Grading scale Imperceptible 8 Perceptible, but not annoying 6 Slightly annoying 4 Annoying Very annoying Table. Modified subjective mean opinion score The resulted scores then were normalized by using the z- scores transform [3]. This transform converts each score into a deviation from the mean score. We got interesting results in the difference in human visual acuity for different types of images. For natural objects, such as an image of the human face, the quality differences can be seen clearly also on low-resolution displays of mobile phones. For the artificial (map) image, the decrease in quality is not so disturbing. For a high-resolution display (CRT), the quality decreases almost linearly. Fig.. z-scores for mobile phone (top), PDA (middle), CRT (bottom). 38
3. SPECTRAL IMAGES 3. Spectral Images Spectral colour imaging is an imaging method, where colour of an object is represented more accurately than in the traditional RGB images. Instead of having only 3 colour components, spectral images have a spectrum connected to each pixel. Spectral imaging is becoming a practical tool in many applications, e.g. in digital commerce, industrial quality control, and digital museum [3]. Comparing to tradition images, spectral images have many advantages. One of the obvious advantages for a consumer is that there is no such issue with white balance that could be seen in digital still cameras. Another side of the coin is that spectral cameras are very expensive and does not suit well for an ordinary customer. Also due to high amount of data, the size of the resulted spectral image is very high. 3. Spectral Image Watermarking In our study we considered a method for watermark embedding into spectral image and studied its properties on a large set of spectral images [4]. By nature, spectral image has 3D form of data. It has two spatial planes and one spectral plane. Thus it was a natural choice to use 3D wavelet domain for watermark embedding and extracting. a a d v,a d v,a d s,d d v,d Fig.3. 3D DWT decomposition Due to high amount of data in spectral images, it is likely that they would be compressed before transmitting. As compression attack we used a PCA-wavelet lossy compression. To reduce the spectral dimension, principal component analysis (PCA) was applied. The compression was achieved by selecting only limited number of principal d h,a d d,a d s,d d v,d d h,d d d,d components to reconstruct the image. The spatial dimension then was compressed using wavelet based SPIHT method. Spectral image is then reconstructed by multiplying the restored principal images by the corresponding principal vectors [5]. The viewing conditions change the perceptual colour of the spectrum. External illumination can be compensated through combining the spectra of the image with the spectrum of the illumination. A set of light sources was used to illuminate the spectral images. Relative spectral radiance factors of the light sources are shown in Fig.4. We evaluated two ways of illumination attack - illumination before watermarking and illumination after watermarking: ) In illumination before watermarking, the original spectral image is multiplied by the illumination vector, and then the result image is watermarked and compressed. The watermark is extracted from the reconstructed image and compared to the original watermark. ) In illumination after watermarking, the original spectral image is watermarked, and then multiplied by the illumination vector, and then the resulting image is compressed. The watermark is extracted from the reconstructed image and compared to the original watermark. We found that in case, we can select values for the watermark strength which gives good quality of the watermarked image and reliable watermark extraction. For the case, we found the value for watermark strength which gives good visual quality of the extracted watermark. The quality of the watermarked image is poor. A proper normalization of illuminated image could improve results. Quality of watermarked image and quality of extracted watermark depends more on illumination then on watermark strength. 39
Radiance.9.8.7.6.5.. 35 4 45 5 55 6 65 7 Radiance.9.8.7.6.5.. 35 4 45 5 55 6 65 7 Radiance.9.8.7.6.5.. 35 4 45 5 55 6 65 7 Fig,4. Relative spectral radiance factors of the light sources. From left to right: A, D65, F. Fig.5. An example of results for watermarking without illumination and compression attack. Original image (left,top), original watermark (left,bottom), watermarked image (right,top), extracted watermark (right, bottom). 4
5. CONCLUSION We have presented here our work for colour image watermarking. We started with traditional RGB images and studied watermarking application for mobile devices. Based on visual quality assessment we can recommend using stronger watermarking for technical images and images that are displayed on small sized displays. Evaluation results shows that with increasing of Image resolution, differences between watermarked image quality on mobile devices and CRT is vanishing, This allows us to use the same watermarking techniques as for desktop application, taking into account hardware limitations. Then we moved forward from traditional images to spectral Images. We presented a technique for watermark embedding and extracting and conducted a set of experiments. A set of illuminators was used to illuminate spectral image. For compression we used PCA-wavelet based method. For Illumination before watermarking we found values for watermark strength that produce good quality images and robust watermarking. For Illumination after watermarking, we found that illumination does effect to quality of watermarked image and extracted watermark much more then watermark strength. Scandinavian Conference on Image Analysis: 3-37. [6] Infotonics Center Joensuu. Spectral Image Database. http://ifc.joensuu.fi. REFERENCES: [] Ingemar J. Cox, Matthew L. Miller, Jeffrey A. Bloom (): Digital Watermarking. Academic Press, San Diego. [] Krasavin K., Parkkinen J., Kaarna A., Jaaskelainen T., (6). Visual quality of watermarking for mobile devices. Journal of SID Vol (No 6). [3] Hordley, S., Finalyson, G., Morovic, P. (4). A multispectral image database and its application to image rendering across illumination. 3rd International Conference on Image and Graphics: 394 397. [4] Krasavin K., Parkkinen J., Kaarna A., Jaaskelainen T. (9). Quality of Reconstructed spectrum for Watermarked Spectral Images Subject to Various Illumination Conditions, th Conference in Advanced Concepts for Intelligent Vision Systems : 57-577. [5] Kaarna, A., Parkkinen, J. (3). Digital Watermarking of Spectral Images with Three-Dimensional Wavelet Transform. 4