ADAPTIVE TECHNIQUES FOR A FAST FREQUENCY DOMAIN MOTION ESTIMATION Yasser Ismail, Mohamed Elgamel, and Magdy Bayoumi The Center for Advanced Computer Studies, University of Louisiana Lafayette {yai6211, ms, mab} @cacs.louisiana.edu ABSTRACT increase the throughput of the encoder but also will increase its complexity. PC technique applied to motion images was Dynamic Block Size Motion Estimation (DBS-ME) and introduced in [7] to tackle these problems by removing both smart Dynamic Early Search Termination (DEST) DCT and IDCT blocks from the inner feedback loop of a techniques are proposed and implemented in this paper. typical video encoder and place the transform block (I) Both of the proposed techniques are combined and applied outside the inner loop as shown in Fig. 1 (b). In this case, a to the conventional phase correlation technique. The Transformed Domain Motion Estimation (TD-ME) will be performance, visual quality and complexity of the proposed used. The main advantage of the structure given in Fig. 1 (b) techniques are compared to that of the original phase is that it increases the throughput due to the removal of both correlation motion estimation (PC-ME) and Full Search DCT and IDCT from the loop, also it decreases the Block Matching (FSBM) techniques. The proposed computational complexity required for motion estimation techniques provide an increase in the encoding quality technique. These advantages make the structure in Fig.1 (b) besides a decrease in the computational complexity of ME more suitable for real time applications [11]. process. Results show that there is approximately 100% of In this paper, we focus on decreasing the computational the stationary blocks decided by the FSBM algorithm are complexity required for the transformed (frequency) domain discovered correctly which consequently reduce the ME process by applying both Dynamic DBS-ME technique computations compared with the original FS and PC which is an accurate and fast MV estimator and DEST techniques. Also it is noted that, DBS-ME technique greatly technique on the conventional PC-ME technique. The decreases the computations required for ME process by proposed encoder in this paper is shown in Fig. 1 (c). It is decreasing the required padding to one or two pixels for noted that, although we increase the complexity of the both the current and the reference blocks. In addition, the encoder in Fig. 1 (c) by adding the inverse transform (T) in motion field of the proposed algorithm gives much lower the feed back loop compared with the encoder given in Fig. 1 entropy than PC-ME which means more reduction in the (b), this complexity is balanced by a huge saving in the transmitted bit rate. computational complexity due to the use of both DEST and 1. INTRODUCTION DBS techniques with the conventional PC technique. In most international video coding standards, such as 2. PHASE CORRELATION TECHNIQUE H.264/AVC [2], Discrete Cosine Transform (DCT) and Phase correlation technique [7] is a frequency domain block based motion estimation are the essential elements to motion estimation method that makes use of the shift achieve spatial and temporal compression respectively. As property of the Fourier transform [9]. According to this seen in Fig. 1 (a), the DCT is located inside the loop of property, a shift in the coordinate frames of two functions is temporal prediction, which also includes an Inverse DCT transformed in the Fourier domain as a linear phase. Phase (IDCT) and a Spatial Domain Motion Estimator (SD-ME) correlation is based on the evaluation of the maximum phase which is usually the Block Match-ME. The coder in this case of the Cross Power Spectrum (CPS) between a pair of co- suffers from both low throughput and high complexity sited rectangular blocks f-l and f, of identical dimensions problems due to the insertion of both DCT and IDCT inside belonging to consecutive frames or fields of a moving the feedback loop. One way to overcome these problems is sequence sampled at t-l, t. Since all functions involved are to design the DCT (or IDCT) to operate at least twice as fast discrete, CPS is circular and it can be carried out as a as the incoming data stream. This solution will not only multiplication in frequency domain using fast Li > > (a) (b) (c) Fig.l1. Coder structures: (a) DCT-ME coder. (b) Frequency domain coder. (c) The proposed Frequency domain coder. 1-4244-1222-6/07/$25.00 ©C2007 IEEE 331 SiPS 2007