1442 C-M.Song et aL Signal Processing:Image Communication 28 (2013)1435-1447 a b 1 88日123 0年123 45678910 4567886 121314151617 12.3146 1819202122 8920212； ●7282 $分mT2万 器7面万2写 456789 4587897 1t234也形 n2314面67 89202T22 8920272 728 728 e 后072了 西50i2写 4667890 256789酒 112134历6 112131456 18920212 8920212 528 728 Fig.4.Comparisons of the motion-compensated 3rd frame produced by our algorithm with those by previous methods at the bit-rate of 128 kbps.The macroblock size is 8 x 8 pixels.(a)Original 3rd frame.(b)1BT,PSNR=21.80.(c)MT-2BT,PSNR=21.97.(d)AQ-2BT.PSNR=22.11.(e)8BT.PSNR=22.79.(f) FQ-2BT.PSNR=22.34. filtering-based one-bit full search [28](1BT).adaptive bits per second).Figs.3 and 4 illustrate the reconstructed quantization-based two-bit full search [23](AQ-2BT). 3rd frames of"Mobile Calendar"sequence using various multithresholding-based two-bit full search [4](MT-2BT) motion estimation algorithms with a block size of 16 x 16 and eight-bit full search with SAD matching metric(8BT). and 8 x 8 pixels,respectively. Since state-of-the-art literatures research only one-or From Fig.3,we can find that the reconstructed frames by two-bit motion estimation,the experimental results of 1BT and MT-2BT algorithms present the most serious ringing our algorithm(FQ-2BT)are not discussed in the case of artifacts at edges.Several numbers on the calendar,e.g.10,17. three-and higher-bit representation in this study. and 24,are very blurred even missing.The result by AQ-2BT algorithm suffers less ringing artifact around number bor- 6.1.Subjective quality comparison ders,but still has number"17"lost and one fake edge at the left side of the calendar.Although the PSNR of FO-2BT is In the first experiment,we compared the subjective 0.14 dB lower than that of 8BT,FQ-2BT and 8BT algorithms quality of decoded frames by our algorithm with those by contribute similar subjective quality which displays clearer representative algorithms at the bit-rate of 128 kbps(kilo numbers than those of 1BT,MT-2BT,and AQ-2BT.filtering-based one-bit full search [28] (1BT), adaptive quantization-based two-bit full search [23] (AQ-2BT), multithresholding-based two-bit full search [4] (MT-2BT), and eight-bit full search with SAD matching metric (8BT). Since state-of-the-art literatures research only one- or two-bit motion estimation, the experimental results of our algorithm (FQ-2BT) are not discussed in the case of three-and higher-bit representation in this study. 6.1. Subjective quality comparison In the first experiment, we compared the subjective quality of decoded frames by our algorithm with those by representative algorithms at the bit-rate of 128 kbps (kilo bits per second). Figs. 3 and 4 illustrate the reconstructed 3rd frames of “Mobile & Calendar” sequence using various motion estimation algorithms with a block size of 16 16 and 8 8 pixels, respectively. From Fig. 3, we can find that the reconstructed frames by 1BT and MT-2BT algorithms present the most serious ringing artifacts at edges. Several numbers on the calendar, e.g. 10, 17, and 24, are very blurred even missing. The result by AQ-2BT algorithm suffers less ringing artifact around number bor￾ders, but still has number “17” lost and one fake edge at the left side of the calendar. Although the PSNR of FQ-2BT is 0.14 dB lower than that of 8BT, FQ-2BT and 8BT algorithms contribute similar subjective quality which displays clearer numbers than those of 1BT, MT-2BT, and AQ-2BT. Fig. 4. Comparisons of the motion-compensated 3rd frame produced by our algorithm with those by previous methods at the bit-rate of 128 kbps. The macroblock size is 8 8 pixels. (a) Original 3rd frame. (b) 1BT, PSNR¼21.80. (c) MT-2BT, PSNR¼21.97. (d) AQ-2BT, PSNR¼22.11. (e) 8BT, PSNR¼22.79. (f) FQ-2BT, PSNR¼22.34. 1442 C.-M. Song et al. / Signal Processing: Image Communication 28 (2013) 1435–1447