1 An Image Refining Method Using Digital Watermark via Vector Quantization *Chin-Chen Chang, **Ju-Yuan Hsiao, and *Po-Wen Lu *Department of Computer Science and Information Engineering National Chung Cheng University, Chiayi, Taiwan 621, R.O.C. FAX: 886-5-2720859 E-mail: {ccc, lpw88}@cs.ccu.edu.tw ** Department of Computer Science and Information Engineering National Changhua University of Education, Changhua 500, Taiwan R.O.C. FAX: 886-4-7211162 E-mail: hsiaojy@cc.ncue.edu.tw Abstract In this paper, we shall propose an image refining method that incorporates the digital watermarking technique and Vector Quantization to recover an image after transmission on a network unclean or tampered by people. In our method, we embed the codevector indices of the original image into the original image itself via the watermarking technique. Then, in the refining phase, the indices are retrieved and employed to detect and recover the supposedly partially damaged image block. The experimental results show that the performance of our method is excellent in suppressing block noise. Keywords: vector quantization, discrete cosine transform, digital watermarking, image refining, denoising 1. Introduction Due to the amazingly fast development of multimedia and network technology, digital images have been more and more widely used in transmitting or spreading messages. However, to transmit an image via network may cause some problems, such as the loss of some part of the image or the tampering of somebody unexpected and unwelcome. In recent years, the digital watermark technique has been pervasively employed to protect the ownership or to identify illegal distributors of the image [4,7,8]. Therefore, the existence of the robust invisible watermark for the image is necessary. The invisible watermark can be embedded without being perceived by human beings; therefore, we can use the invisible watermark to assist the influenced image refining process. Our method incorporates digital watermarking and builds it onto the Vector Quantization (VQ) technique that is popular for image compression [1-3,5-6,9-12]. In the following paragraph, there will be a brief introduction of the concept of VQ. In the process of VQ on an image, the source image is divided into blocks or vectors [12]. A block or vector is composed of L pixel values. In other words, the vector is an L-dimensional vector. The vector qauantizer can encode the vectors of the source image into indices in a codebook or decode the indices into codevectors of the codebook. At both the encoding and decoding processes of vector quantizer, we have a set of L-dimensional codevectors called the codebook, and each codevector is assigned to with a binary index. In the encoding process, an input vector is compared to each codevector in the codebook in order to find the codevector that is closest to the input one. The index of the closest codevector is the quantized value of the input vector. In contrast, those codevectors corresponding to the indices of the source image are retrieved to recover the original image in the decoding process. In our method, we first compress the original image by VQ. Then, we embed the indices of the codevectors of the original image into the coefficients of the DCT blocks [8] via the watermarking technique [12]. After the watermarked image is manipulated, the image is called an influenced image. In order to refine the influenced image, we first retrieve the embedded indices, and then the indices are employed to help detect and recover a noised block. The rest of this paper is organized as follows. Sections 2 and 3 are the watermark (index) embedding process and extracting process, respectively. The refining procedure is described in Section 4. Finally, the experiments and conclusions are shown in Sections 5 and 6, respectively. 2. Watermark (Index) Embedding Process In our method, the source image with MM pixels in it is divided into NN blocks or vectors of the size or dimension L=(MM)/(NN). These NN vectors are encoded as NN indices where each block or vector corresponds to an index. In the index embedding procedure, to reduce the effect of the damage to the original image, each bit of index is embedded in the DCT block coefficient of t Proceedings of the 2005 11th International Conference on Parallel and Distributed Systems (ICPADS'05) 0-7695-2281-5/05 $20.00 © 2005 IEEE