"HEVC inter-frame prediction paper" P1.1 fast inter-frame prediction algorithm based on motion feature in HEVC

based on the motion characteristics HEVC fast inter-frame prediction algorithm /fast inter-frame prediction Algorithm for HEVC Based on Motion Features

"HEVC standard Introduction, HEVC Frame prediction paper Notes" series blog, catalogue see: Http://www.cnblogs.com/DwyaneTalk/p/5711333.html

Journal of Shanghai University (Natural Science Edition) Vol. 19, section 3.

The current CU's motion characteristics are judged by using the square and mean of the difference of the luminance pixel value of the encoded CU in the current depth CU and the time domain corresponding position . Belongs to a Class A algorithm

ASSD:Average Sum of the Square difference

Core formula:

2N represents the current the dimensions of Cu,cucur (i,j), and Cucol (i,j) represent the luminance pixel values in the current and time domain corresponding to the position of the CU block (i,j) respectively.

The larger the ASSD, the greater the difference between the adjacent cu in the time domain indicates the more intense the motion.

ASSDX represents the value of ASSD that has been encoded (should be encoded in the current frame)when the best depth in CU is X (0~3) . That is , ASSDX only consider CU with the best coding depth of x , and other parameters such as ASSD.

ASSDNotX represents the value of ASSD that has been encoded in the best depth of the CU, greater than X (x+1~3) . So assdnotx contains assdx+1~assd4.

Thskip indicates that the code is encoded The mean value of the ASSD of all cu in the CU with the best depth greater than X. If the current CU block depth x ASSD is greater than Thskip, indicating that the current CU block motion is very intense, so there is no need for depth x under the various PU prediction mode to calculate and judge, jump directly to the next depth x+1 for a smaller size cu judgment. In the formula B is the adjustment parameter, the text takes 1.2. The larger the value, the greater the corresponding Thskip.

Thstop indicates that the code is encoded The mean value of the ASSD of all cu in the CU with the best depth of x. If the ASSD under current CU block depth x is less than thstop, the average of all blocks with an optimal depth of x in the current frame's encoded CU is smoothed, as it can be assumed that the current CU block is in a flat area. For the CU in the flat zone , the deeper CU division judgment is not continued. Where A is the adjustment parameter, the text takes 0.8. a smaller, the smaller the corresponding thstop

Core judgment:

by definition, the general Thskip is greater than thstop, and when ASSD is in between, the image is in a moderate motion zone. The algorithm acceleration optimization is mainly aimed at the motion region and the gentle movement area, and it can't accelerate the moderate movement. It is necessary to control the size of Thskip and thstop reasonably. If [Thstop,Thskip] interval is too large, the acceleration effect is not good, if [Thstop,Thskip] interval is too small, resulting in a large number of judgements for the movement of vigorous and gentle movement of the region, prone to miscarriage, affecting image quality. Therefore, it is necessary to reasonably control the value of a, B, thskip > 2*thstop are generally indicated.

Algorithm Flow:

The main steps are as follows : (It is indicated that the non-symmetrical pu mode is not considered)

1, calculates the current  cu   assd,  and according to the encoded  cu  Information (thskip and thstop) ,  Judging the motion characteristics of the current  cu 

2, if the current  cu  in the active area and its depth is less than  maxdepth,   Skips the calculation of the current  cu  all predictive modes ,  into the next depth   cu  encoding, because in this case the current depth of all pu mode is not optimal, so do not try ;  if the current  cu in the active region and the coding depth is equal to  maxdepth,  the   Skip, 2nx2n, nx2n, 2nxn  and  nxn  and other modes to make predictions ,  and calculate rate distortion generation value ; 

3, if the current cu is in a moderate motion area and the coding depth is less than MaxDepth, then the other modes except Skip and NXN mode to perform rate distortion generation value calculation ; if the current cu is in a moderate motion area and the encoding depth is equal to M Axdepth, the current skip mode is skipped, and the Rate Distortion generation value of 2nx2n, nx2n, 2NxN and NXN mode is calculated.

4, if the current cu in the movement of the smooth zone , then no longer continue to divide, because at this time the depth of the CU is considered sufficient, do not need further division .

Experimental results:

"HEVC inter-frame prediction paper" P1.1 fast inter-frame prediction algorithm based on motion feature in HEVC