A. Skodras, "Fast Discrete Cosine Transform Pruning", IEEE Trans. Sig. Proc., July 1994  Home/Search   Document Not in Database   Summary   Related Articles   Check

....reductions but may suffer larger degradations. We can set up more thresholds to perform DCT more adaptively, for example, calculating lower 2x2 or 4x4 DCT coefficients. To calculate part of DCT coefficients, DCT pruning algorithms can be used. Several papers have addressed the DCT pruning issue [17][18] and have demonstrated that computations can be reduced compared to calculating all 8x8 DCT coefficients. 4. Simulation Results We use public domain H.263 softwares (both TMN5 [19] and TMN8 [20] for simulation. TMN5 uses full search on motion estimation and TMN8 uses fast search. Various ....

A. Skodras, Fast Discrete Cosine Transform Pruning, IEEE Transaction on Signal Processing, vol. 42, no. 7, pp. 1833-1837, July 1994.

....degradation will be graceful. To be on the conservative side, we will use 99 probability in our following simulations. 3 Computation Reduction of DCT To calculate the DCT coefficients before the EOB, DCT pruning algorithms can be used. Several papers have addressed the DCT pruning issue [1] 2][5] and have demonstrated that computations 5 can be saved if we only need to calculate part of the DCT coefficients (in both 1 D and 2 D cases) In this paper, we propose an adaptive method based on the quantization parameter Q for a simplified two level DCT pruning algorithm. The DCT pruning ....

....defined by Fu Cu fm ( cos( 2 21 16 0 (4 1) where Cu( 12, for u = 0 ; Cu( 1 otherwise. Several fast algorithms have been proposed for the standard DCT [3] 4] If we only need the first four coefficients, several pruning algorithms such as Wang [1] and Skodras [5] (for the 1 D case) and Christopoulos [2] for the 2 D case) can be used. For example, using Wang s pruning algorithm, calculating the first 4 coefficients of an 8 point 1 D DCT needs 22 additions and 8 multiplications. In Eq 4 1, the first cosine term is cos( up 16 , the second term is cos( ....

A. Skodras, "Fast Discrete Cosine Transform Pruning," IEEE Transaction on Signal Processing, vol. 42, no. 7, pp. 1833-1837, July 1994.

....size N due to the downsampling operation, i.e. N = Mk; k 2 Z (48) M is the decimation factor and Z denotes any positive integer) the choice of N is much more flexible compared with the PIPO algorithms. The other advantage of the SIPO approaches is in the computation of the pruning DCT [15]. In the DCT based signal compression algorithms, the most useful information of the signal is kept in the low frequency DCT components. Therefore, retaining only N 0 N coefficients is sufficient for the lossy data compression. Although the pruning DCT can be computed from the PIPO DCT ....

....of the signal is kept in the low frequency DCT components. Therefore, retaining only N 0 N coefficients is sufficient for the lossy data compression. Although the pruning DCT can be computed from the PIPO DCT architecture by removing the unnecessary data paths and computational operators [15], the global communication is still the major drawback for its implementation as N increases. On the contrary, the SIPO architecture in [3] and our low power design can be readily applied to the pruning DCT by simply implementing the first N 0 DCT modules for the computation of the first N 0 DCT ....

A. N. Skodras, "Fast discrete cosine transform pruning," IEEE Trans. Signal Processing, vol. 42, pp. 1833--1837, July 1994.

No context found.

A.N. Skodras, "Fast discrete cosine transform pruning", IEEE Trans. on Signal Processing. Accepted.

.... [3] Furthermore, in [3] a vector radix pruning method is described which allows immediate computation of the lowest N 0 xN 0 out of NxN DCT components (where both N and N 0 are powers of two) This vector radix pruning method is compared with the row column pruning technique proposed recently in [11,13]. 2.0 Performance comparison The computational complexity of the direct 2D FCT algorithm is given by NxN mN = 3 4 2 and a NxN mN N N = 3 2 2 2 2 , where NxN and a NxN denote the number of multiplications and additions respectively [2,3] Compared to the 2D FCT computed via the ....

....N=2 m , N 0 =2 m0 ) It s pruning property makes it a useful tool in applications in which only a predetermined number of points has to be calculated as in many adaptive compression schemes [4] This results in large computational gain. In Fig. 2. the execution times for the row column pruning [11,13] approach and the vector radix pruning are given. The size NxN of the input data is chosen 256x256 and the computed N 0 xN 0 DCT components are the low frequency 16x16, 32x32, 64x64, 128x128 and 256x256 points. It can be observed that in all computer architectures the vector radix approach is much ....

A. N. Skodras, "Fast discrete cosine transform pruning", IEEE Trans. on Signal Processing. Accepted.

....arithmetic) This observation leads us to examine whether the computation of the DCT may be simplified so that not all elements of O have to be computed. In essence, in the case of the fast implementations of the DCT, some of the butterfly calculations that are required can be altogether avoided [9]. This speeds up the processing and produces longer run lengths of zero coefficients as well; the latter property typically results in better compression. The above introduce us to the concept of the pruning level of a DCT computation. A pruning level of p denotes that the upper left p p ....

A.N. Skodras, "Fast Discrete Cosine Transform Pruning," IEEE Trans. on Signal Processing, Vol. 42, No. 7, 1994, pp. 1833-1837.

....2. I]Z#[adl# gVe]#d[#)m)# 9#9 H#;8I###C b a a b = cos( p , C a a = cos( p ) 3. 1 Pruned Flow Graph Analysis At each stage s, s=0,1, m 1, there are B s blocks or bunches of 2x2 butterflies, and b s 2x2 butterflies per block (Figure 2) where B s 2 2m 2s 2 = and b s 2 2s = [4,8]. Let the positive integers q and r be the quotient and the remainder of (N0 2 b s ) respectively, that is N0 2 q b s r = 6) A number of complete butterflies (n cb ) and a number of incomplete butterflies (n ib ) has to be computed at each stage. The exact number of complete and incomplete ....

....approach. It is evident that better results will be obtained for triangular pruning using the recursive pruning technique described in section 4. The N 0 xN 0 points can be calculated by means of a 1D pruning algorithm applied in the row column manner. Using the 1D pruning algorithm proposed in [8,10] the number of multiplications ( N rc 0 ) and additions (a N rc 0 ) equals to [4] N rc m N N N = 12) a N rc m N m N N N = 13) A comparison of these equations with eq. 10) and eq. 11) corresponding to the vector radix FCT ....

[Article contains additional citation context not shown here]

A.N. Skodras (1994): Fast discrete cosine transform pruning", IEEE Trans. on Signal Processing. 42(7).

No context found.

A. Skodras, "Fast Discrete Cosine Transform Pruning", IEEE Trans. Sig. Proc., July 1994

No context found.

A. N. Skodras, "Fast discrete cosine transform pruning," IEEE Trans. Signal Processing, vol. 42, pp. 1833--1837, July 1994.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC