CN101577825B - Interactive quantized noise calculating method in compressed video super-resolution - Google Patents

Abstract

The invention discloses an interactive quantized noise calculating method in compressed video super-resolution. The method includes the following steps of: first counting the appearance probability ofDCT coefficients before quantization before a coding end does quantized operation to the DCT coefficients of a video frame and calculating Laplace parameter capable of representing the distribution o f the DCT coefficients before quantization; then writing the distribution parameter of the DCT coefficients before quantization of an obtained image block in a Data-user field reserved in a code stream to be sent to a decoding end through coding; and finally obtaining the distribution parameter of the DCT coefficients before quantization in the code stream from the decoding end, calculating and obtaining the quantized noise according to the distribution probability density of the DCT coefficients before quantization and coefficients after quantization and consequently obtaining a final high-resolution image in a super-resolution algorithm. The calculating method is applicable to the compressed video super-resolution algorithm interacting between the coding end and the decoding end, improves the accuracy of quantized noise by obtaining the distribution parameter of the DCT coefficients before quantization at the coding end and providing the distribution parameter for the decoding end tocalculate the quantized noise.

Description

Interactive quantized noise calculating method in the compressed video super-resolution

Technical field

The present invention relates to interactive quantized noise calculating method in a kind of compressed video super-resolution, belong to field of multimedia communication.

Background technology

In video monitoring, for the needs that transmit and store, the resolution of image is often less, and the open market main flow is CIF (352*288) resolution, and the resolution during playback video often is difficult to the identification monitored object.The way of lifting resolution commonly used is to adopt better watch-dog, but such investment is big and existing equipment can't be reused, thereby under the prerequisite that does not increase new watch-dog, the resolution that promotes monitor video by software seems very urgent and important.In traditional super-resolution rebuilding algorithm, at first reason such as optical dimming, motion blur, down-sampling and the noise to image deterioration carried out mathematical modeling, then and the inversion operation by matrix obtain high-resolution image, the model that therefore degrades accurately has important effect to super-resolution rebuilding.In compressed video, quantification is the main cause that causes information loss in the compression process, just need obtain the distribution character that quantizes preceding DCT coefficient in order to obtain accurate quantization noise model, existing algorithm is with DCT coefficient distributed model parameter or be assumed to be preset parameter or it is carried out iterative as variable, accuracy is not high, but also has problems such as computation complexity height.Conventional model meets even distribution to generally adopting the preceding DCT domain coefficient of quantification to distribute to the modeling of quantizing noise in the method for reconstructing of compressed video, studies show that, the DCT coefficient distributes and satisfies laplacian distribution before quantizing, traditional like this can not accurately express quantizing noise at DCT coefficient distributed model before quantizing in the compressed video super resolution reconstruction algorithm, therefore can influence effect and the quality that the supercompression video super-resolution is rebuild.On the other hand, simple at decoding end rebuilding super resolution image, there are two problems:

(1) quantize before the DCT coefficient distribute and meet laplacian distribution, but decoding end has only the back of quantification DCT coefficient, calculates quantizing noise and need introduce new variable and characterize the laplacian distribution parameter, obtains the value of this parameter by iteration optimization, has increased computing cost; (2) quantizing noise characterizes in the DCT territory and calculates, and needs to introduce the frequency domain constraint in the super-resolution rebuilding algorithm, the computation complexity height, and convergence is slow.

Summary of the invention

In order to solve above-mentioned the deficiencies in the prior art, the present invention proposes a kind of computational methods of compressed video super-resolution interactive quantized noise, DCT coefficient distributed model parameter before the video coding end calculates quantification, then it is delivered to decoding end, and decoding end quantizes preceding DCT coefficient distribution probability density and quantizes back DCT coefficient by image block to obtain more accurate quantizing noise, reconstructs better video quality.

Technical scheme of the present invention may further comprise the steps:

(1) before coding side frame of video DCT domain coefficient carries out quantization operation, the number of times that DCT coefficient before quantizing occurs is added up, obtain the probability that each coefficient occurs, the object of statistical operation is a DCT coefficient before quantizing in the image block, and the size of image block is the integral multiple of discrete cosine transform block size in the coding; Calculate the probability that quantizes preceding DCT coefficient appearance according to following formula:

$P\left(y_i\right)=\frac{C\left(y_i\right)}{\underset{1}{\overset{n}{\mathrm{\Σ}}}C\left(y_i\right)}---\left(1\right)$

Wherein, DCT coefficient before the different quantification in the i presentation video piece, n are the total numbers of DCT coefficient before the quantification in the image block, the number of times that the DCT coefficient occurred before C (y) expression quantized, and P (y) is corresponding probability.

(2) will quantize preceding DCT coefficient and corresponding probability and bring Laplce's probability distribution formula calculating acquisition distributed constant into; Get the laplacian distribution parameter according to following formula:

$P\left(y\right)=\frac{\mathrm{\λ}}{2}{e}^{-\mathrm{\λ}\left|y\right|}---\left(2\right)$

Wherein, the distribution probability of DCT coefficient before P (y) expression quantizes, y is DCT coefficient before quantizing, λ represents the laplacian distribution parameter, in practical operation, DCT coefficient 0 before quantizing is added up, and the distributed constant that obtains with this value is as the parameter of token image piece, λ=2P (0).

(3) the laplacian distribution parameter value that will obtain image block writes the Data_user field of reserving in the code stream, sends to decoding end by coding side; At coding side, the parameter that the DCT coefficient distributed before the sign that obtains quantized writes the Data_User field in the code stream, according to the image block sequence arrangement of correspondence, sends to decoding end by code stream when coding.

(4) in obtaining the Data_user field of code stream, decoding end extracts DCT coefficient laplacian distribution parameter before the quantification of correspondence image piece.

(5) calculate the quantizing noise value of each quantization parameter according to DCT coefficient distribution probability densitometer before quantizing quantification that back DCT coefficient and coding side transmit; Calculate the value of quantizing noise according to following formula:

$k=\frac{1}{q}{\∫}_{y_q-\frac{q}{2}}^{y_q+\frac{q}{2}}{\left(y{-y}_q\right)}^{2}p\left(y\right)\mathrm{dy}---\left(3\right)$

Wherein, q is a quantization step, and y is DCT coefficient before quantizing, y _qFor quantizing back DCT coefficient, k represents the value of quantizing noise.

The present invention and prior art contrast, effect is positive and tangible: the present invention is applicable to the bigger situation of quantizing noise influence in the compressed video, obtain to quantize preceding DCT coefficient distributed constant by coding side and be transferred to decoding end, directly calculate quantizing noise value accurately, coding and decoding improves the model that obtains of quantizing noise in the super-resolution rebuilding of compressed video alternately.

Description of drawings

Fig. 1 is an original image.

Fig. 2 is the image that reconstruction in 30 o'clock obtains for quantization step.

Fig. 3 is that the quantizing noise compensation of coefficient calculations obtains reconstructed image before utilizing decoding end to quantize.

Embodiment

Below the present invention is further illustrated by specific embodiment:

For calculation procedure is described, from Fig. 1, get a discrete cosine transform block arbitrarily computational process is described, we get first in the present embodiment, and its image block matrix is I1, then has:

$I1=\left[\begin{array}{cccccccc}137& 137& 138& 136& 138& 129& 138& 134\\ 137& 137& 136& 136& 138& 129& 139& 134\\ 137& 138& 137& 136& 139& 129& 137& 134\\ 137& 137& 138& 135& 137& 129& 138& 134\\ 137& 138& 137& 136& 138& 129& 139& 133\\ 140& 140& 131& 130& 136& 133& 132& 134\\ 134& 135& 141& 132& 134& 137& 132& 127\\ 133& 133& 129& 133& 131& 133& 129& 131\end{array}\right]$

Numeric representation pixel value wherein.I1 is carried out dct transform, obtains matrix I2:

$I2=\left[\begin{array}{cccccccc}1079& 10& 1& -1& 1& 0& -8& 9\\ 9& -1& 1& -5& 3& 3& -7& 7\\ -6& -1& -2& 1& -1& 0& 4& -4\\ 3& 2& 2& 2& -1& -2& 0& 1\\ -1& -2& 0& -2& 2& 2& -2& 0\\ 1& 2& -3& -1& -4& 0& 4& 2\\ -1& 0& 4& 2& 4& -4& -2& -3\\ 2& 1& -4& -2& -2& 2& 0& 2\end{array}\right]$

DCT coefficient before wherein numeric representation quantizes.In the quantization operation, we get quantization step q=30, obtain quantizing back DCT coefficient matrix I3:

$I3=\left[\begin{array}{cccccccc}36& 0& 0& 0& 0& 0& 0& 0\\ 0& 0& 0& 0& 0& 0& 0& 0\\ 0& 0& 0& 0& 0& 0& 0& 0\\ 0& 0& 0& 0& 0& 0& 0& 0\\ 0& 0& 0& 0& 0& 0& 0& 0\\ 0& 0& 0& 0& 0& 0& 0& 0\\ 0& 0& 0& 0& 0& 0& 0& 0\\ 0& 0& 0& 0& 0& 0& 0& 0\end{array}\right]$

(1) use following formula to calculate and quantize the probability that preceding DCT coefficient occurs:

$P\left(y_i\right)=\frac{C\left(y_i\right)}{\underset{1}{\overset{n}{\mathrm{\Σ}}}C\left(y_i\right)}---\left(1\right)$

Wherein, DCT coefficient before the different quantification in the i presentation video piece, n is the total number of the coefficient in the image block, the number of times that the DCT coefficient occurred before C (y) expression quantized, P (y) is corresponding probability.According to analyzing as can be known in an embodiment n=63 (first coefficient is represented DC coefficient among the I2, does not participate in statistics).

(2) calculate the laplacian distribution parameter according to following formula:

$P\left(y\right)=\frac{\mathrm{\λ}}{2}{e}^{-\mathrm{\λ}\left|y\right|}---\left(2\right)$

The distribution probability of DCT coefficient before P (y) expression quantizes, y is DCT coefficient before quantizing, λ represents the laplacian distribution parameter, probability density when using the DCT coefficient to be 0 value is found the solution the λ value, make y=0: λ=2P (0) can obtain λ=2P (0)=8/63 from I2.

(3) for each image block to obtaining Laplce's parameter distribution of a correspondence, at coding side, these parameters are write Data_User field in the code stream, when coding,, send to decoding end by code stream according to the image block sequence arrangement of correspondence.

(4) the Laplce's parameter in decoding end reception code stream, the λ value of acquisition respective image piece.

(5) calculate the value of single quantizing noise according to following formula:

$k=\frac{1}{q}{\∫}_{y_q-\frac{q}{2}}^{y_q+\frac{q}{2}}{\left(y{-y}_q\right)}^{2}p\left(y\right)\mathrm{dy}---\left(3\right)$

Wherein, q is a quantization step, and y is DCT coefficient before quantizing, y _qFor quantizing back DCT coefficient, k represents the value of quantizing noise.We find from I3: quantization step is 30 o'clock, and it is 0 that DCT quantizes the back coefficient, and can draw: q=30, y _q=0, $p\left(y\right)=\frac{\mathrm{\λ}}{2}{e}^{-\mathrm{\λ}\left|y\right|},$ λ＝8/63。Pass through integral and calculating $k=\frac{1}{30}\underset{-15}{\overset{15}{\∫}}{y}^{3}p\left(y\right)\mathrm{dy}=1.230246879827472.$ The quantizing noise that calculates can compensate in the image reconstruction and go.

Calculate the image subjective quality that quantizes loss in compensation below by the quantizing noise that relates among example explanation the present invention:

Fig. 1 is original image lena.bmp, and Fig. 2 is the image that reconstruction in 30 o'clock obtains for quantization step, and Fig. 3 obtains reconstructed image for utilizing the preceding coefficient calculations of decoding end quantification to obtain the quantizing noise compensation.By observing, can find that calculating quantizing noise alternately by the encoding and decoding end can obtain subjective picture quality preferably in compressed image is rebuild.

Claims (5)

1. interactive quantized noise calculating method in the compressed video super-resolution is characterized in that may further comprise the steps:

(1) before coding side frame of video DCT domain coefficient carries out quantization operation, the number of times that DCT coefficient before quantizing occurs is added up, obtain each and quantize the probability that preceding DCT coefficient occurs, the object of statistical operation is a DCT coefficient before quantizing in the image block, and the size of image block is the integral multiple of discrete cosine transform block size in the coding;

(2) will quantize preceding DCT coefficient and corresponding probability and bring Laplce's probability distribution formula calculating acquisition distributed constant into;

(3) the image block laplacian distribution parameter value that obtains is write the Data_user field of reserving in the code stream, send to decoding end by coding side;

(4) in obtaining the Data_user field of code stream, decoding end extracts DCT coefficient laplacian distribution parameter before the quantification of correspondence image piece;

(5) according to quantizing the quantizing noise value that DCT coefficient distribution probability before the quantification that back DCT coefficient and coding side transmit calculates each quantization parameter.

2. according to interactive quantized noise calculating method in the described compressed video super-resolution of claim 1, it is characterized in that: calculate the probability that quantizes preceding DCT coefficient appearance according to following formula in the step (1):

$P\left(y_i\right)=\frac{C\left(y_i\right)}{\underset{1}{\overset{n}{\mathrm{\Σ}}}C\left(y_i\right)}---\left(1\right)$

Wherein, y _iCoefficient value before quantizing in the presentation video DCT coefficient block, DCT coefficient before the different quantification in the i presentation video piece, n are the total numbers of DCT coefficient before the quantification in the image block, C (y _i) expression quantize before the number of times that occurs of DCT coefficient, P (y _i) be corresponding probability.

3. according to interactive quantized noise calculating method in the described compressed video super-resolution of claim 1, it is characterized in that: step gets the laplacian distribution parameter according to following formula in (2):

$P\left(y\right)=\frac{\mathrm{\λ}}{2}{e}^{-\mathrm{\λ}\left|y\right|}---\left(2\right)$

Wherein, the distribution probability of DCT coefficient before P (y) expression quantizes, y is DCT coefficient before quantizing, λ represents the laplacian distribution parameter, in practical operation, to DCT coefficient value before quantizing is that 0 number of times that occurs is added up, and is worth the parameter of the distributed constant of acquisition as the token image piece with this, gets λ=2P (0).

4. according to interactive quantized noise calculating method in the described compressed video super-resolution of claim 1, it is characterized in that: in the step (3), at coding side, the parameter that the DCT coefficient distributed before the sign that obtains quantized writes the Data_User field in the code stream, when coding,, send to decoding end by code stream according to the image block sequence arrangement of correspondence.

5. according to interactive quantized noise calculating method in the described compressed video super-resolution of claim 1, it is characterized in that: the value of calculating quantizing noise in the step (5) according to following formula:

$k=\frac{1}{q}{\∫}_{y_q-\frac{q}{2}}^{y_q+\frac{q}{2}}{(y-y_q)}^{2}P\left(y\right)\mathrm{dy}---\left(3\right)$

Wherein, q is a quantization step, and y is DCT coefficient before quantizing, y _qFor quantizing back DCT coefficient, k represents the value of quantizing noise, the distribution probability of DCT coefficient before P (y) expression quantizes.

Images