**1. Introduction**

18 Will-be-set-by-IN-TECH

[20] Loaiciga, H. [1989]. An optimization approach for groundwater quality monitoring

[21] Mahar, P. & Datta, B. [1997]. Optimal monitoring network and ground-water-pollution source identification, *Journal of water resources planning and management* 123(4): 199–207. [22] Mahar, P. S. & Datta, B. [2000]. Identification of pollution sources in transient

[23] Mahar, P. S. & Datta, B. [2001]. Optimal identification of ground-water pollution sources and parameter estimation, *Journal of Water Resources Planning and Management-Asce*

[24] Mahinthakumar, G. & Sayeed, M. [2005]. Hybrid genetic algorithmlocal search methods for solving groundwater source identification inverse problems, *Journal of water*

[25] Michalak, A. M. & Kitanidis, P. K. [2004]. Estimation of historical groundwater contaminant distribution using the adjoint state method applied to geostatistical inverse

[26] Pinder, G. [2009]. Optimal search strategy for the definition of a dnapl source, *Technical*

[27] Singh, R. & Datta, B. [2006]. Identification of groundwater pollution sources using ga-based linked simulation optimization model, *Journal of hydrologic engineering* 11: 101. [28] Singh, R. & Datta, B. [2007]. Artificial neural network modeling for identification of unknown pollution sources in groundwater with partially missing concentration

[29] Singh, R., Datta, B. & Jain, A. [2004]. Identification of unknown groundwater pollution sources using artificial neural networks, *Journal of water resources planning and management*

[30] Singh, R. M. & Datta, B. [2004]. Groundwater pollution source identification and simultaneous parameter estimation using pattern matching by artificial neural network,

[31] Sun, A. Y., Painter, S. L. & Wittmeyer, G. W. [2006a]. A constrained robust least squares approach for contaminant release history identification, *Water Resour. Res.* 42(4): W04414. [32] Sun, A. Y., Painter, S. L. & Wittmeyer, G. W. [2006b]. A robust approach for iterative contaminant source location and release history recovery, *Journal of Contaminant*

[33] Sun, N.-Z. [1994]. *Inverse Problems in Groundwater Modeling*, Kluwer Academic

[34] Zheng, C., Wang, P. P., Zheng, C. & Wang, P. P. [1999]. Mt3dms: A modular three-dimensional multi-species transport model for simulation of advection, dispersion, and chemical reactions of contaminants in ground-water systems. documentation and user's guide, *Contract Report SERDP-99-1, U.S. Army Engineer Research and Development*.

*report*, Defense Technical Information Center, United States of America.

observation data, *Water Resources Management* 21: 557–572.

network design, *Water Resources Research* 25(8): 1771–1782.

*resources planning and management* 131: 45.

modeling, *Water Resour. Res.* 40(8): W08302.

*Environmental Forensics* 5(3): 143–153.

*Hydrology* 88(3-4): 181–196.

Publishers, pp. 12–37.

127(1): 20–29.

130: 506.

groundwater systems, *Water Resources Management* 14(3): 209–227.

The promising video coding standard, H.264/AVC [1], is developed by the Join Video Team of ITU-T Video Coding Experts Group (VCEG) and ISO/IEC Moving Picture Experts Group (MPEG). By utilizing several new techniques, such as advanced intra predictions, variable block size ME, integer transformation, in-loop deblocking filter, H.264/AVC has achieved significant compression gain compared with previous video coding standards. It is now widely applied to many types of visual services, for example Digital Multimedia Broadcasting, Mobile Phone, and High Definition (HD) video delivery. In the near future, holography video and Super-HD video are expected to hit consumer market. These kinds of large sized video contents require higher coding efficiency while keeping the encoder complexity within an acceptable level. Therefore, new techniques are needed to reduce the computational complexity so that various real time video encoder and delivery services for the large sized video contents could be feasible.

In particular, Block-Matching Motion Estimation (BMME) with Full Search (FS) algorithm [2] is the main computational burden in H.264/AVC due to exhaustively search all possible blocks within the search window using Lagrangian multiplier. Although FS algorithm can obtain the optimum motion vector (MV) in most cases, it consumes more than 80% of the total computational complexity. Thus, a fast and efficient motion estimation algorithm is required for H.264/AVC. Recently, two major approaches were researched to overcome this problem. One employs fast mode decision algorithms to skip unnecessary block modes in variable block checking process [3, 4]. The other one utilizes Fast Motion Estimation (FME) searching algorithms to reduce unnecessary search points [5-11].

Various algorithms have been proposed to reduce search points for FME Search algorithm. Motion adaptive search (MAS) [5] utilized the motion activity information to adjust the search strategy. In Variable Step Search (VSS) algorithm [6], motion search range is

© 2012 Shi et al., licensee InTech. This is an open access chapter distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/3.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. © 2012 Shi et al., licensee InTech. This is a paper distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/3.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

determined by using the degree of correlation between neighbouring motion vectors. A Multi-Path Search (MPS) algorithm [7] has been proposed, in which all the eight neighbours around the origin of the search window were performed to find candidate points. This algorithm has good rate-distortion performance, but its computational complexity reduction is limited. To tackle this drawback, the directional gradient descent search (DGDS) algorithm [8] is developed. It searched on the error surface in eight directions by using directional gradient descent. The search patterns in each stage depend on the minima found in eight directions and thus the global minimum can be traced more efficiently.

Simulated Annealing for Fast Motion Estimation Algorithm in H.264/AVC 177

step 1 step 2 step 3 step 4-1 step 4-2 step 5-1 step 5-2

*d1*

*d22*

*d24*

*d2*

vector existing in the same directional region is at least 75%. Based on this statement, PIDS algorithm exploited the predicted MV direction to decide the intensive search direction. Thus, the intensive-direction search and coarse-direction search are selected adaptively for different regions. One example of search process of PIDS is depicted in Figure 2. As the uneven search pattern is changing according to the predicted motion vector of each block, it performs more precisely than UMHexagonS and achieves more computation reduction.

*d7 d4*

However, the PIDS algorithm's adaptive intensive search selection is limited in directional regions. With fixed number of search points in each direction, it cannot adjust the search range for different motion scenes. In study [11], a statistic analysis of MV distribution was carried out. A large number of global minima occupy near the search centre especially at the zero MV (0, 0) with a certain percentage of optimal MVs outside the radius of 10 pels. It indicated that most predicted and optimal MVs have high locality correlation. Meanwhile, some irregular MVs can hardly be well predicted due to poor correlation. In this chapter, direction and distance correlation between predicted MV and optimal MV are investigated MV correlation statistics information is calculated for each frame as its motion characteristic. With this information, the intensive and coarse search regions are adaptively changed for each block. The Simulated Annealing concept [12, 13] is employed to control searching process and to adaptively choose the intensive search region. After this Introduction the


*d16 d19*

Section 2 statistical analyse MV direction and distance correlation characteristic. The blockmatching motion estimation is described in this section. Section 3 gives an overview of simulated annealing and simulated quenching algorithm. Based upon analyses, the proposed SAAS algorithm is presented in section 4. The experimental results are given and

**Figure 2.** Search pattern of PIDS, example of intensive search in *d1*




0

5

10

*d10*

*d13*

15

chapter is organized into five more sections as follows.

illustrated in section 5. Finally, section 6 draws the final conclusion.

The hybrid multi-hexagon-grid search (UMHexagonS) algorithm [9] was adopted in H.264/AVC reference software JM as its significant reduce the computational complexity with only little degradation in rate-distortion performance. UMHexagonS takes advantage of four kinds MV predictions to decided initial search point, i.e. the Median Prediction (MP), the Uplayer Prediction (UP), the Corresponding-block Prediction (CP) and the Neighbouring Reference-picture Prediction (NRP). After selecting the best initial point, it employs the unsymmetrical-cross search pattern and uneven-hexagon-grid search pattern, which are shown in Figure 1 as step2 and setp3-2. In these uneven search patterns, the number of horizontal search points is more than that of vertical points. This is mainly based on a common assumption, that the movement in the horizontal direction is higher than that in vertical direction. However, motion characteristic in each video sequence is unique. Also, the characteristic may change with the time. Therefore, with this horizontal-heavy pattern, UMHexagonS would lose accuracy and waste searching power.

**Figure 1.** Search process of UMHexagonS algorithm

Predictive Intensive Direction searching (PIDS) algorithm was proposed in [10] to solve the problem caused by uneven search patterns by using a adaptive searching pattern. In PIDS algorithm, the correlation between predicted motion vector and optimal motion vector was investigated. The study revealed that the probability of predicted and optimum motion vector existing in the same directional region is at least 75%. Based on this statement, PIDS algorithm exploited the predicted MV direction to decide the intensive search direction. Thus, the intensive-direction search and coarse-direction search are selected adaptively for different regions. One example of search process of PIDS is depicted in Figure 2. As the uneven search pattern is changing according to the predicted motion vector of each block, it performs more precisely than UMHexagonS and achieves more computation reduction.

**Figure 2.** Search pattern of PIDS, example of intensive search in *d1*

176 Simulated Annealing – Single and Multiple Objective Problems

determined by using the degree of correlation between neighbouring motion vectors. A Multi-Path Search (MPS) algorithm [7] has been proposed, in which all the eight neighbours around the origin of the search window were performed to find candidate points. This algorithm has good rate-distortion performance, but its computational complexity reduction is limited. To tackle this drawback, the directional gradient descent search (DGDS) algorithm [8] is developed. It searched on the error surface in eight directions by using directional gradient descent. The search patterns in each stage depend on the minima found

The hybrid multi-hexagon-grid search (UMHexagonS) algorithm [9] was adopted in H.264/AVC reference software JM as its significant reduce the computational complexity with only little degradation in rate-distortion performance. UMHexagonS takes advantage of four kinds MV predictions to decided initial search point, i.e. the Median Prediction (MP), the Uplayer Prediction (UP), the Corresponding-block Prediction (CP) and the Neighbouring Reference-picture Prediction (NRP). After selecting the best initial point, it employs the unsymmetrical-cross search pattern and uneven-hexagon-grid search pattern, which are shown in Figure 1 as step2 and setp3-2. In these uneven search patterns, the number of horizontal search points is more than that of vertical points. This is mainly based on a common assumption, that the movement in the horizontal direction is higher than that in vertical direction. However, motion characteristic in each video sequence is unique. Also, the characteristic may change with the time. Therefore, with this horizontal-heavy pattern,


Predictive Intensive Direction searching (PIDS) algorithm was proposed in [10] to solve the problem caused by uneven search patterns by using a adaptive searching pattern. In PIDS algorithm, the correlation between predicted motion vector and optimal motion vector was investigated. The study revealed that the probability of predicted and optimum motion

step2 step3-1 step3-2 step4-1 step4-2

in eight directions and thus the global minimum can be traced more efficiently.

UMHexagonS would lose accuracy and waste searching power.


**Figure 1.** Search process of UMHexagonS algorithm



0

5

10

15

However, the PIDS algorithm's adaptive intensive search selection is limited in directional regions. With fixed number of search points in each direction, it cannot adjust the search range for different motion scenes. In study [11], a statistic analysis of MV distribution was carried out. A large number of global minima occupy near the search centre especially at the zero MV (0, 0) with a certain percentage of optimal MVs outside the radius of 10 pels. It indicated that most predicted and optimal MVs have high locality correlation. Meanwhile, some irregular MVs can hardly be well predicted due to poor correlation. In this chapter, direction and distance correlation between predicted MV and optimal MV are investigated MV correlation statistics information is calculated for each frame as its motion characteristic. With this information, the intensive and coarse search regions are adaptively changed for each block. The Simulated Annealing concept [12, 13] is employed to control searching process and to adaptively choose the intensive search region. After this Introduction the chapter is organized into five more sections as follows.

Section 2 statistical analyse MV direction and distance correlation characteristic. The blockmatching motion estimation is described in this section. Section 3 gives an overview of simulated annealing and simulated quenching algorithm. Based upon analyses, the proposed SAAS algorithm is presented in section 4. The experimental results are given and illustrated in section 5. Finally, section 6 draws the final conclusion.
