DocumentCode :
146812
Title :
Motion vector dependent filter strength computation for high performance H.264 de-blocking loop filter
Author :
Nandan, Nibedita ; Mody, Mihir ; Sanghvi, Hetul
Author_Institution :
Texas Instrum. Ltd., Bangalore, India
fYear :
2014
fDate :
3-5 April 2014
Firstpage :
404
Lastpage :
408
Abstract :
Block based video codec inevitably induce blocking artifacts due to discrete quantization approach. Blocking artifacts impairs subjective picture quality and coding efficiency. De-blocking loop filter reduces visually objectionable artifacts associated with the coding of macroblocks (MB) and their constituent sub-blocks. Filter operation is adaptive in nature and strength of filter (strong/weak) depends on quantum of introduced blocking artifact, measured through Boundary strength (BS). Out of 5 BS values (0,1,2,3,4), computation cost of motion vector (MV) and reference picture dependent BS=1 and BS=0 is in range of 21- 12 cycles for hardware based solutions. To meet 4k@60 fps performance LPF needs to finish each mb filter in 100 cycles and BS=1/0 computation will take minimum 12% of cycle budget. The paper proposes a novel approach to use hardware accelerator of entropy coding (ECD) to pre-calculate BS=1/0 value for LPF. The proposed scheme results in significant silicon area savings as well as reduction in processing cycles. The approach enables reaching performance goal of 4K@60 fps in TI´s video engine (IVAHD2).
Keywords :
entropy codes; filters; video codecs; video coding; H.264 de-blocking loop filter; IVAHD2; boundary strength; coding efficiency; deblocking loop filter; discrete quantization; entropy coding; hardware accelerator; macroblocks coding; motion vector dependent filter strength computation; video codec; video engine; Vectors; 4K; BS; De-blocking loop filter; H.264; LPF; NF1; boundary strength; motion vector;
fLanguage :
English
Publisher :
ieee
Conference_Titel :
Communications and Signal Processing (ICCSP), 2014 International Conference on
Conference_Location :
Melmaruvathur
Print_ISBN :
978-1-4799-3357-0
Type :
conf
DOI :
10.1109/ICCSP.2014.6949872
Filename :
6949872
Link To Document :
بازگشت