H.264/AVC 利用可變區塊大小(variable block size) 的動作預估 (motion estimation)和編碼率失真最佳化(rate-distortion optimized) 的區塊模式選擇 (macroblock mode decision) 達到極高壓縮效果,但是暴力的區塊模式檢查過程花費了整個壓縮過程大部分的計算量。為了要降低這部份的計算量,許多論文提出了加速模式選擇的演算法,卻忽略了前景、背景和區塊模式間的關係。 本篇論文提出了基於背景模型的模式選擇演算法:我們首先利用本質影像技術產生背景資訊,根據背景資訊,每個宏塊 (macroblock) 將會背歸類到三種前景、背景的宏塊類型當中,每種類型將會指定使用不同的候選區塊模式。區塊模式的編碼率失真值 (rate-distortion cost) 將成為剔除不必要的模式檢查的比較基準。另外,背景資訊也會用來加速動作預估。本論文也提出一個依QP值不同而改變的臨界值,當做提早結束區塊模式選擇的標準。 實驗結果顯示,平均而言本論文可以加速50%的編碼時間,而只造成少部分的失真率下降以及編碼率提高。另外,我們也試著探究影像中真正的運動區域,同時也維持了前景部分的影像品質。
H.264/AVC utilizes variable block size motion estimation and rate distortion optimized mode decision to achieve high coding efficiency. However, the exhaustive mode checking procedure takes up most part of the computational load. To reduce the computational complexity, many works focus on expediting the mode decision process but are not aware of the relationship between the foreground/background and the chosen mode. In this thesis, we propose a background-modeling-based fast mode decision algorithm which will first generate the background information using intrinsic image technique. According to the background information, macroblocks are classified into three foreground/background macroblock types which will be assigned different candidate modes. Rate-distortion cost will be the comparison to skip unnecessary block mode checking, and with the assistance of background information, fast motion estimation mechanism is proposed. We also use a QP dependent threshold value as an early termination yardstick. The experimental results demonstrate that the proposed algorithm reduces, on average, 48% of the entire encoding time with negligible PSNR drop and bit-rate increase. Moreover, we endeavor to find out the true motion in the video sequences, while retaining the foreground quality.