ON THE ADAPTIVE 3-DIMENSIONAL TRANSFORM CODING TECHNIQUES FOR MOVING IMAGES by JONG

In this paper, based on the 3-dimensional discrete cosine transform(DCT) for removing the temporal correlation, an adaptive 3-D transform coding technique is proposed. Because of the nonstationary nature of the image data, the energy distribution in a 3-D DCT block varies along the vertical, horizontal and temporal directions. Thus, adaptive schemes, such as the 3-D classi cation, the classi ed linear scanning technique and the VLC table selection scheme, are used to take into account of local variations. Also, in our approach, a hybrid technique, which adaptively combines relatively simple inter-frame coding with intra-frame coding, is presented. Through intensive computer simulations, the performance of the proposed The authors are with Signal Processing Lab., Dept. of Control and Instrumentation Eng., Seoul National University, Shinlim-Dong, Kwanak-Ku, Seoul 151-742, KOREA, and also a liated with the Automatic Systems Research Institute(ASRI), Seoul National University, Seoul, 151-742, KOREA. Phone number: (822) 880-7309 , Fax number: (822) 885-6620 , E-mail address: [email protected] To be addressed all the correspondence. 3-D transform coding technique is evaluated on the several well-known moving sequences. The results show that, especially for moving sequences containing slow or moderate motion, the proposed technique provides an improved performance over the scheme with motion compensation [1] at rates above 0.5 b/pixel(bpp), and a good visual quality of the reconstructed images is obtained. Thus, the proposed 3-D transform coding technique is believed to be a good candidate for the digital VCR, since the motion compensation is not required in the proposed 3-D coding technique. I . Introduction I . Introduction An e ective coding scheme for moving image sequences exploits both the temporal and spatial redundancies. Hence the motion compensated transform coding(MCTC) is one of the most popular techniques currently used for encoding the moving images [2]. However, the computational complexity required for the motion compensation, such as motion estimation, is very high [3{5], especially for a high quality moving images. Also, sometimes so called the motion artifacts are observed if the employedmotion information is inaccurate. An alternative approach to decorrelate the temporal redundancy, without the use of motion compensation technique, is a 3-dimensional(3-D) transform coding technique [6{11]. The performance of the 3-D transform coding may be inferior to the 2-dimensional(2-D) MCTC technique for low bit rates, since the motion compensation is proven to be very e ective in reducing the temporal redundancy with relatively small transmission cost. The increase in time delay and the required storage capacity with increasing number of frames are another disadvantages of the 3-D transform coding technique. However, the 3-D transform coding technique is not dependent on the previously coded frames. So, no refreshing frames need to be sent to the decoder and the reconstruction at arbitrary frame position can be implemented easily. The 3-D transform coding employing a simple zonal sampling was rst reported by Roese, Pratt and Robinson [6]. However, due to the nonstationary nature of the image data, the information content of the transformed samples changes from one block to the other, and this a ects the performance of the encoder signi cantly. The performance of the 3-D transform coder can be improved by considering adaptive methods [8{11]. Hence, in this paper, we propose a new adaptive 3-D transform coding technique for encoding a moving image sequence. Among many di erent types of unitary transforms, we consider the discrete cosine transform(DCT) in our approach, since it has been shown that the theoretical mean square error coding performance of the DCT is nearly equivalent to that obtained by the optimal Karhunen-Lo eve transform for highly correlated data [12]. The