Signal Processing Image Processing Paper ID 1338 Reducing Prediction Error in Watermark Retrieval Process by Multiple Non-linear Filtering Narong Mettripun, Nwe Ni Hlaing, Thitiporn Pramoun, Thumrongrat Amornraksa Multimedia Communication Laboratory, Computer Engineering Department, Faculty of Engineering, King Mongkut's University of Technology Thunburi, 126 Pracha-uthit Rd., Bangmod, Thungkru, Bangkok, 10140, Thailand mettripun _ n@hotmail.com, nwenihlaing28@gmail.com Abstract- This paper presents an improving method used in the watermark retrieval process of digital watermarking for color image based on the modiication of image pixels. Based on the prior knowledge of the embedded watermark pattern, some errors from the watermark bit prediction process can be reduced simply by changing the predicted bit value in accordance with all predicted bits value within the prediction area. To achieve this, we simply implement a non-linear ilter with different characteristics to the predicted watermark bits, and output the result as the retrieved watermark. With the proposed method, the performance in both accuracy of the retrieved watermark and robustness of the embedded watermark can be improved. The experimental results show the improvements in term of NC at the equivalent PSNR ater implementing our retrieval method with the existing watermarking scheme, compared to the previous methods. The improved robustness of the embedded watermark against common image processing based attacks and geometrical attacks are also evaluated and compared. I. INTRODUCTION In digital world, digital media such as image and video can be easily copied and distributed. In order to protect their copyright, digital watermarking have been introduced and developed. Actually, those methods help discourage unauthorized people rom violating digital property of the other ones. Over 10 years, image watermarking has still been an attractive issue for many researchers around the world because a large numbers of digital image are continually distributed in the cyber space through networks. Principally, image watermarking consists of two main processes which are watermark embedding and retrieval processes. The irst one is to embed the owner information into an original host image known as watermark, while the second one is to retrieve the embedded watermark rom the watermarked image. There are two image domains for embedding a watermark, namely, the requency domain and spatial domain [1] and [2]. In requency domain, we can embed a watermark into coeficients of the transformed image, while in spatial domain; we can embed a watermark into a host image directly by modiying the image pixels. There is a large amount of previous work on image watermarking. For example, in the requency domain based approach, Wang et al. [3] considered the perceptual invisibility and robustness for digital watermarking and proposed the method of watermark embedding in the discrete wavelet transform (DWT) coeicients for ownership veriication requires. Suhail et al. [4] developed digital watermarking algorithm based on discrete cosine transform (DCT) coeicients and image segmentation. Satio The 8th Electrical Engineering/ Electronics, Computer, Telecommunications and Information Technology (ECTI) Association of Thailand - Conference 2011 and Kino [5] proposed the digital watermarking method in color images based on wavelet and HVS (Human Visual system). The concept of HVS can be implemented with image watermarking in order to strengthen the watermark signal without any noticeable change to human eyes. Although the requency domain based watermarking methods are well resistant against various images processing based attacks, it is not strong enough against geometrical attack, e.g. image scaling. Moreover, if the watermark bits are embedded too much in the requency domain, the image quality will be signiicantly degraded [6]. For the spatial domain based approach, Kutter et al. [7] proposed the method to embed a large number of watermark bits into a color image in blue channel by modiying the image pixel using either additive or subtractive, depending on the watermark bit, and proportional to the luminance of the embedding pixel. Their approach was proved to be robust against various types of attacks especially geometrical attacks. The robustness of their watermarking method was improved later by set of techniques, proposed by T. Amoraksa et al. [8]. Those techniques were balancing the watermark bits around the embedding pixels, tuning the strength of embedding watermark in accordance with the nearby luminance, and reducing the bias in the prediction of the original image pixel rom the surrounding watermarked image pixels. The authors also demonstrated how to embed a watermark image into a color image having the same resolution i.e. by embedding mXn watermark bits into mXn color image pixels as well. However, the precision in the original pixel prediction process was degraded when applied with an image having high variation of pixel values. Subsequently, to improve performance of the above watermark retrieval process, Z. L. Aung et al. [9] proposed the watermark retrieval method based on non-linear iltering. Accordingly, the watermark signal in their method was considered as a noise, and thus can be removed by one of an eicient noise reduction techniques i.e. non-linear iltering. Recently, N. N. Hlaing et al. [10] proposed a method to reduce the prediction error rom the watermark bit prediction process. Based on the knowledge of embedded watermark patten, this prediction eror can be reduced by changing the output bit value in accordance with the voting result rom the predicted bits value within the prediction area. However, their method implemented a ixed patten and repeated the process with this patten until obtaining satisying watermark accuracy. Obviously, this method introduced unnecessary complexity to the watermarking system. Page 987