A ZeroValue Prediction Technique for Fast DCT Computation

1
A Zero-Value Prediction Technique for Fast DCT
Computation

• Y. Nishida, K. Inoue, and V. Moshnyaga

2
Introduction

• Discrete Cosine Transform (DCT) is the most
  commonly
• used transform for video coding (JPEG, MPEG
  H26x etc).

Problem

• DCT is a very computationally expensive task.

• DCT requires a measurable amount of energy
• and many operations.

Reducing the number of operations and energy
consumption of DCT becomes important.
2
3
Our Proposal
A Zero-Value Prediction Technique for Fast DCT
Computation
Advantage

• Reduces the number of DCT and quantization
  
• operations.
• Reduces energy consumption.
• Increases DCT operation speed.

Disadvantage

• The picture quality deteriorates.

3
4
The feature of DCT
The high frequency components at the DCT output
are concentrated nearby zero value.
DCT
Input data (Pixel Level)
Output data (DCT coefficient)
4
5
The feature of Quantization
The probability of elements in the encoding block
to become zero after quantization is very high.
Quantization
Input data (DCT coefficient)
Quantization output data
In the case Quantization step 16
5
6
Zero-Value Prediction Technique
When the DCT output N zeros, we predict the
remaining result is also zero without actual DCT
computing.
In the case with N2
The detected string of zeros
The predicted value
Starting point of zero-value prediction
Reducing operations for 35 DCT coefficients is
possible.
6
7
A Zero-Value Prediction Technique Example (1/2)
DCT coefficient
Conventional Technique
A Zero-Value Prediction technique
The prediction error increases but most of the
DCT coefficients will be zeros after
quantization.
7
8
A Zero-Value Prediction Technique Example (2/2)
Quantization output data
Conventional Technique
A Zero-Value Prediction technique
Prediction mistake
Prediction mistake
When a picture is decoded, the prediction error
degrades image quality
8
9
Evaluation
Using the MPEG software from MPEG Software
Simulation Group

• 2D-DCT algorithm use addition 1024 and
  multiplication 1024 times
• per 1 block.
• 2D-DCT use partitionable method (1D-DCT1D-DCT).
• Raster scan order of DCT computation.

Video benchmarks
missa
carpohne
foreman
salesman
QCIF
QCIF
QCIF
CIF
150frame
382frame
298rame
300rame
9
10
PSNR as a function of N
?
?
carphone
foreman
missa
salesman

?
70
60
50
PSNR (dB)
40
30
20
10
0
3
1
2
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
N

• The PSNR is low when N is small.
• The PSNR does not change a lot after N9.

10
11
Operation reduction rate as a function of N(DCT)
?
?
carphone
foreman
missa
salesman

Operation reduction rate
?
0.5
0.4
0.3
0.2
0.1
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
N

• The reduction rate is high when N is small.

11
12
Operation reduction rate as a function of
N(quantization)
?
?
carphone
foreman
missa
salesman

?
Operation reduction rate
1.0
0.8
0.6
0.4
0.2
0
3
1
2
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
N

• The reduction rate is high when N is small
• (similarly to the DCT operation).

12
13
Example of reconstructed picture for N9
Assumption to accept N9 as a tradeoff for both
picture quality and the reduction rate.
The example of a picture
Original method
Zero-value prediction(N9)
The PNSR decreases by 1.07dB.
48.75dB
47.67dB
13
14
Conclusion
A Zero-Value Prediction Technique for Fast DCT
Computation

• When N9, the number of computations is reduced
  by 29 for DCT and by 59 for quantization.
• The average PSNR value degreases by -1.6dB.

Future work

• Estimating the effect of the proposed technique
  on energy consumption.

14
15
Comparison of the algorithm
The number of operations per 1block.
Normal addition 1024 and multiplication 1024
times. Chen addition 416 and multiplication
224 times. FFT addition 464 and
multiplication 176 times.
15