Adaptive Image Compression Using Local Pattern Information Source

Adaptive Image Compression Using Local Pattern Information Source: Pattern Recognition Letters 23 (2002) 1837 -1845 Author: Feng Pan Speaker: Hsien-Wen Tseng Date: Oct. 17, 2002

Outline w Block pattern classification w Adaptive quantization w Adaptive Zigzag scan w Experimental results w Conclusions 2

Block pattern classification 3

Block pattern classification w HE = [ (v+1)2×F 2(u, v)]1/2 w VE = [ (u+1)2×F 2(u, v)]1/2 w DE = [ (u+1) ×(v+1)×F 2(u, v)]1/2 w HV = arctan (VE/HE) 4

Block pattern classification 1 2 3 4 5 6 7 8 1 434 0 0 0 0 2 -57 0 0 0 0 3 63 0 0 0 0 4 -78 0 0 0 0 5 118 0 0 0 0 6 -395 0 0 0 0 7 -154 0 0 0 0 8 38 0 0 0 0 HE = [12×(-57)2+ 12×(63)2+ 12×(-78)2+…+ 12×(38)2]1/2 = 456 VE = [22×(-57)2+ 32×(63)2+ 42×(-78)2+…+ 82×(38)2]1/2 = 2713 DE = [1× 2× (-57)2+ 1× 3× (63)2+ 1× 4× (-78)2+…+ 1× 8× (38)2]1/2 = 1107 HV = arctan(VE/HE) = 80 o 5

Block pattern classification w Horizontal edge: HV 32. 5 o w Vertical edge: HV 57. 5 o w Diagonal edge: 32. 5 o < HV < 57. 5 o, DE > 200 w None-edge: 32. 5 o < HV < 57. 5 o, DE 200 6

Block pattern classification 7

Adaptive quantization QF(u, v) = QHVS(u, v)× 0. 7 QH(u, v) = QHVS(u, v)×[0. 09×(v-7)+1. 3] QV(u, v) = QHVS(u, v)×[0. 09×(u-7)+1. 3] QD(u, v) = QHVS(u, v)×[0. 09×(|u-v|-7)+1. 3] 8

Adaptive quantization 9

Adaptive zigzag scan Horizontal edge Vertical edge 10

Experimental results 11

Conclusions w Adaptive quantization - reduce the quantization error of the important coefficients. w Adaptive zigzag scan – reduce the size of run -length coding. w 2 penalty bits per block are needed. 12