CN103338363B - Video compress perceptual coding system based on measurement field block sort and method - Google Patents

Abstract

A kind of video compress perceptual coding system based on measurement field block sort and method, system includes splits' positions sampling module, the block sort module of measurement field, sample rate distribution module, hits distribution and double measurement module and reconstructed module.Input video block is divided into the sub-block of non-overlapping copies equal in magnitude by coding side, measures each sub-block with identical random measurement matrix and obtains measurement field signal；The block sort module of measurement field obtains measurement field characteristic image using measurement field signal as the row vector of matrix, set up the correlative relationship model of measurement field signal and frequency-region signal, with this relational model for according to utilizing measurement field characteristic image to carry out the judgement of block classification, sample rate is distributed to different types of video block, hits distribution and double measurement module determine measurement number and calculation matrix according to sample rate and video block, utilize calculation matrix that video signal is remeasured to obtain measuring signal, measurement signal is transferred to reconstructed module reconstruct.

Description

Video compress perceptual coding system based on measurement field block sort and method

Technical field

The invention belongs to technical field of video coding, relate generally to video coding system based on compressed sensing and method, a kind of video compress perceptual coding system based on measurement field block sort and method, for video coding system.

Background technology

In H.264 standard video coder system, estimation is used to remove the temporal correlation of video signal, use dct transform to remove spatial coherence, in fact further study show that, either yet suffer from stronger dependency between intra block or the DCT coefficient of interframe block, this dependency is strengthened along with the closing on of locus of block, and this dependency is referred to as frequency domain and is correlated with by us.

Signal collecting device based on compressed sensing utilize compressed sensing technology that the optical signal of simulation directly changes into compression after digital signal, therefore in coding side original, pixel domain video signal non-availability.The traditional estimation in video encoding standard, motion compensation, Predicting Technique etc. cannot be applicable to video coding systems based on compressed sensing, therefore research is needed to analyze method based on the signal correlation in measurement field, to improve the compression efficiency of whole system at coding side.Or assuming that coding side can obtain original pixel domain digital video signal, in pixel domain, directly analyze signal correlation, and this has obviously run counter to compressive sensing theory；Or introducing restructing algorithm at coding side, the pixel-domain video signal obtained based on reconstruct analyzes its dependency, substantially increases coding side complexity.Therefore, coding side utilize low complexity algorithm obtain picture signal pixel domain feature the most difficult.

nullDocument " ZhirongGao，ChengyiXiong，ChengZhouandHanxinWang，CompressiveSamplingwithCoefficientsRandomPermutationsforImageCompression，InternationalConferenceonMultimediaandSignalProcessing(CMSP)，pp.321-324，2011 " mention different image blocks in, at transform domain, there is different degree of rarefications and complexity，Smooth block is relative to texture block less summation about non-zero DCT coefficients corresponding with edge block，The amplitude change ratio of texture block DCT coefficient is shallower，The amplitude change of edge block DCT coefficient is more violent.That is frequency-region signal feature can effectively reflect the pixel domain feature of image.

From analysis above, in video coding system based on compressed sensing, by the relation between research measurement field signal correlation and corresponding frequency domain signal correlation, directly utilize measurement field signal and realize the correlation research of frequency-region signal, there is very important researching value.

In video coding system based on compressed sensing, generally frame of video is carried out piecemeal process to reduce the complexity of memory space and reconstruct, and all of video block or image block use identical processing method, and have ignored the feature of dissimilar video block itself.In the application of most of video communications, people's major concern is subjective quality rather than the objective quality of image.Additionally consider the feature of the visual system of human eye, the sensitivity difference to the zones of different human eye of image, utilize block sort method to divide the image into different regions, then different to different area applications Processing Algorithm.

Existing block sort method generally utilizes the characteristic distributions of the mean variance in pixel domain or frequency-region signal as classification foundation, and the coding side pixel domain signal of video coding system of based on compressed sensing and frequency-region signal non-availability, existing block sort method is unavailable, therefore in video coding system based on compressed sensing, study the block sort method in measurement field, carry out different process for different types of video block and will assist in the distortion performance of raising system.

Summary of the invention

It is an object of the invention to the shortcoming overcoming above-mentioned prior art, a kind of video compress perceptual coding system based on measurement field block sort and method are proposed, feature for dissimilar video block carries out compression in various degree to different types of video block, to improve the distortion performance of whole system.

For achieving the above object, the invention provides a kind of video compress perceptual coding system based on measurement field block sort, including splits' positions sampling module, the block sort module of measurement field, sample rate distribution module, hits distribution and double measurement module and reconstructed module；Input video block is divided into the sub-block of non-overlapping copies equal in magnitude by splits' positions sampling module, measures each sub-block with identical random measurement matrix and obtains measurement field signal；nullThe block sort module of measurement field is by extracting measurement field characteristic image module、Measurement field is constituted with frequency domain relational model module and judging module，The measurement field signal of sub-block splits' positions sampling obtained obtains measurement field characteristic image as the row vector of matrix，Measurement field establishes the linear approximate relationship of measurement field signal and the dependency of corresponding frequency-region signal to frequency domain relational model module，Judging module utilizes measurement field characteristic image to carry out the judgement of block classification with the correlative relationship model of measurement field Yu frequency-region signal for theoretical foundation，It is specially the Cross-covariance of computation and measurement characteristic of field image，And show that by correlative relationship model the variance of element of Cross-covariance is for weighing the feature of video block，Size according to variance yields realizes the type judgement of video block，By judging module, the video block of input is judged to edge block、Texture block or smooth block；Then system is for court verdict and the feature of dissimilar video block itself, the sample rate different to the distribution of different types of video block；Hits distribution and double measurement module determine measurement number and calculation matrix according to sample rate and video block sizes, utilize calculation matrix that video block re-starts measurement and obtain measurement field signal, finally measurement field signal are transferred to reconstructed module and are reconstructed.

Video block is classified by prior art by obtaining original, pixel domain signal in coding side reconstruct, greatly add the complexity of coding side, or assume that coding side has original, pixel domain signal, run counter to the original intention of compressive sensing theory, and the present invention establishes measurement field signal and discrete cosine transform (DiscreteCosineTransformation, DCT) the correlative relationship model of territory signal, and utilize this model to achieve block sort in measurement field, carry out sample rate distribution for judgement type, improve the distortion performance of whole system.

The realization of the present invention also resides in the described dependency setting up measurement field and frequency domain relational model cross covariance two signals of measurement, according to video signal coefficient characteristic distributions after discrete cosine transform, establish the linear approximate relationship model between the dependency of measurement field signal and corresponding frequency domain signal correlation.

The present invention utilizes measurement field signal to achieve the correlation analysis of corresponding DCT domain signal, and classifying for the video block in measurement field provides theoretical foundation.

The present invention also provides for a kind of video compress perceptual coding method based on measurement field block sort, and the system that the method is suitable for is the video compress perceptual coding system based on measurement field block sort that the present invention proposes, and the method comprises the steps:

Step 1. is by input picture block or video block x_bIt is divided into the sub-block of non-overlapping copies equal in magnitude, it is assumed that be divided into four sub-blocks x_b1、x_b2、x_b3、x_b4, corresponding DCT domain signal is θ₁、θ₂、θ₃、θ₄, the gaussian random calculation matrix of each sub-block formed objects is measured respectively and obtains measurement field signal y₁, y₂, y₃, y₄, measured value is the column vector of m × 1；

Step 2. using the measurement field signal of sub-block as matrix row vector build measurement field characteristic image y=[y₁, y₂, y₃, y₄]^T；

Step 3. sets up the correlative relationship model of measurement field signal and frequency-region signal:

$C_y=\frac{n}{m}{C}_{\mathrm{\θ}}$

Wherein, C_yFor the Cross-covariance in measurement field, n is the pixel count of video block, and m is for measuring number, C_θFor the Cross-covariance in DCT domain；

Each element of measurement field covariance matrix is sought variance by step 4., the relational model of step 3 obtain:

$\mathrm{var}\left(C_y\right)={\left(\frac{n}{m}\right)}^2\·\mathrm{var}\left(C_{\mathrm{\θ}}\right)$

Wherein, variance, C are asked in var () expression_yFor the Cross-covariance of measurement field characteristic image, n is the pixel count of video block, and m is for measuring number, C_θFor the Cross-covariance in DCT domain.

The type of video block is made decisions by step 5. by following criterion, according to the variance size of the Cross-covariance element of measurement field characteristic image video block is judged to smooth block, edge block or texture block:

Wherein if be " if " the meaning, T₁, T₂For empirical value.

The video block type that step 6. exports according to judging module carries out sample rate distribution, the distribution of sample rate meets edge block ＞ texture block ＞ smooth block, size according to sample rate and video block determines hits and calculation matrix, then utilizes this calculation matrix that current video block re-starts measurement and obtains measurement field signal y_b。

Step 7. is by measurement field signal y_bIt is transferred to reconstructed module, utilizes compressed sensing restructing algorithm that current video block is reconstructed.

The dependency of the element representation sub-block of Cross-covariance and the variance of sub-block itself, the most compared with prior art, the measurement field block sort method that the present invention proposes not only allows for the global statistics characteristic of video block signal, considers the local distribution characteristic of video block signal the most to a certain extent.

It is as follows with the establishment step of the correlative relationship model of frequency-region signal that the realization of the present invention also resides in step 3 measurement field signal:

1) according to the character of calculation matrix, the relational model of measurement field covariance matrix and frequency-region signal is set up:

$C_y\≈\frac{1}{m}\mathrm{\θ}{\mathrm{\θ}}^T$

Wherein, C_yFor the Cross-covariance of measurement field characteristic image, m is for measuring number, and θ is DCT domain sample matrix；

2) according to video signal in the characteristic distributions of DCT domain, the relational model of frequency domain Cross-covariance and frequency-region signal is set up:

$C_{\mathrm{\θ}}=\frac{1}{n}\mathrm{\θ}{\mathrm{\θ}}^T$

Wherein, C_θFor the Cross-covariance in DCT domain, n is the pixel count of video block, and θ is DCT domain sample matrix；

3) by step 1) and 2) relational model set up the correlative relationship model of measurement field signal and frequency-region signal:

$C_y\≈\frac{n}{m}{C}_{\mathrm{\θ}}$

Wherein, C_yFor the Cross-covariance of measurement field characteristic image, n is the pixel count of video block, and m is for measuring number, C_θFor the Cross-covariance in DCT domain.

The present invention devises video compress perceptual coding system based on measurement field block sort, it is provided that video compress perceptual coding method based on measurement field block sort, compared with prior art has the advantage that

(1) present invention achieves the video block classification in measurement field, in coding side pixel domain signal non-availability or the classification remaining able to realize video block under conditions of not being reconstructed, feature according to dissimilar video block carries out compression in various degree, improves the distortion performance of system；

(2) present invention establishes the correlative relationship model of measurement field signal and frequency-region signal, utilizes measurement field signal to achieve the correlation analysis of frequency-region signal；

(3) the block sort method of measurement field that the present invention realizes not only allows for the statistical property of current video block, and considers the local distribution characteristic of video block to a certain extent, improves the accuracy of block sort.

Accompanying drawing explanation

Fig. 1 is the composition block diagram of present invention video compress based on measurement field block sort perceptual coding system.

Fig. 2 is the method flow schematic diagram of the present invention.

Fig. 3 is the experimental result of the measurement field block sort of the lena image of the present invention, wherein Fig. 3 (a) is original image, and Fig. 3 (b), 3 (c) and 3 (d) are respectively the Fig. 3 (a) smooth block, edge block and texture block after block sort.

Detailed description of the invention

1 and Fig. 2 the present invention is described in further detail below in conjunction with the accompanying drawings:

Step 1. splits' positions is sampled

By input picture block or video block x_bIt is divided into the sub-block of non-overlapping copies equal in magnitude, the gaussian random calculation matrix of each sub-block formed objects is measured respectively and obtains measured value.The specific embodiment of the present invention is that the video block intending sampling is further divided into four sub-blocks x_b1、x_b2、x_b3、x_b4, corresponding DCT domain signal is θ₁、θ₂、θ₃、θ₄, the gaussian random calculation matrix of each sub-block formed objects to be sampled respectively, sample rate is MR=0.5, respectively obtains measurement field signal y₁, y₂, y₃, y₄, measurement field signal is the column vector of m × 1；

Step 2. extracts measurement field characteristic image

Using the measurement field signal of sub-block as matrix row vector build measurement field characteristic image

$y={[y_1,y_2,y_3,y_4]}^T={\left[\begin{array}{cccc}{y}_{11}& y_{21}& y_{31}& y_{41}\\ y_{12}& y_{22}& y_{32}& y_{42}\\ \·& \·& \·& \\ \·& \·& \·& \\ \·& \·& \·& \\ y_{1m}& y_{2m}& y_{3m}& y_{4m}\end{array}\right]}^T=\left[\begin{array}{cccc}{y}_{11}& y_{12}& \·\·\·& y_{1m}\\ y_{21}& y_{22}& \·\·\·& y_{2m}\\ \·& \·& & \·\\ \·& \·& & \·\\ \·& \·& & \·\\ y_{41}& y_{42}& \·\·\·& y_{4m}\end{array}\right];$

Step 3. seeks the Cross-covariance of measurement field characteristic image

3a) set up the correlative relationship model of measurement field signal and frequency-region signal:

$C_y=\frac{n}{m}{C}_{\mathrm{\θ}}$

Wherein, C_yFor the Cross-covariance of measurement field characteristic image, n is the pixel count of video block, and m is for measuring number, C_θFor the Cross-covariance in DCT domain；

3b) seeking the Cross-covariance of measurement field characteristic image, the specific embodiment of the present invention obtains:

$C_y=\left[\begin{array}{cccc}{C}_{11}& C_{12}& C_{13}& C_{14}\\ C_{21}& C_{22}& C_{23}& C_{24}\\ C_{31}& C_{32}& C_{33}& C_{34}\\ C_{41}& C_{42}& C_{43}& C_{44}\end{array}\right]$

Wherein, matrix element C_ij=Cov(y_i, y_j)=E{ [y_i-E(y_i)][y_j-E(y_j)] it is second order mixing central moment, and 1≤i≤4,1≤j≤4.

Diagonal entry C_ij(i=j)=E{ [y_i-E(y_i)]²Represent measurement field signal y_iVariance, off diagonal element C_ij(i ≠ j) represents sub-block y_iAnd y_jBetween dependency.

Each element of measurement field Cross-covariance is sought variance, by step 3a by step 4.) relational model obtain:

$\mathrm{var}\left(C_y\right)={\left(\frac{n}{m}\right)}^2\·\mathrm{var}\left(C_{\mathrm{\θ}}\right)$

Wherein, variance, C are asked in var () expression_yFor the Cross-covariance of measurement field characteristic image, n is the pixel count of video block, and m is for measuring number, C_θFor the Cross-covariance in DCT domain.

The type of video block is made decisions by step 5. by following criterion, according to the variance size of the Cross-covariance element of measurement field characteristic image video block is judged to smooth block, edge block or texture block:

Wherein if be " if " the meaning, T₁, T₂For empirical value.

The specific embodiment of the present invention takes T₁=3 × 10⁶, T₂=2 × 10⁷。

The distribution of step 6. sample rate and double measurement

Video block type according to judging module output carries out sample rate distribution, the distribution of sample rate meets edge block ＞ texture block ＞ smooth block, size according to sample rate and video block determines hits and calculation matrix, then utilizes this calculation matrix that current video block re-starts measurement and obtains measurement field signal.Video block is divided in the specific embodiment of the present invention edge block, texture block and smooth block, and each video block sample rate is taken as 0.5,0.4 and 0.3 respectively.

Measurement field signal is transferred to reconstructed module by step 7., utilizes compressed sensing restructing algorithm to be reconstructed current video block.The specific embodiment of the present invention use base follow the trail of (BasicPursuit, BP) algorithm is reconstructed, use dictionary that adjacent reconstructing video frame constructs as sparse base (see document LiuHaixiao, SongBin, QinHao, QiuZhiliang, ADictionaryGenerationSchemeforBlock-BasedCompressedVideo Sensing, ProceedingsofICSPCC, Sep2011).

The 3 couples of present invention block sort effect in measurement field is described further below in conjunction with the accompanying drawings:

Realize the simulated conditions of Fig. 3: hardware environment: CPUAMDSempron3000+, 1.8GHz, 512MB internal memory；Software environment: WindowsXP, MatlabR2010b；Image block size: 8 × 8；Reference picture: lena；Resolution: 256 × 256.

As can be seen from Figure 3, the measurement field block sort method that the present invention proposes preferably can carry out block sort to image or frame of video, and image 3 (a) substantially can accurate be divided into smooth block 3 (b), edge block 3 (c) and texture block 3 (d).

Below in conjunction with table 1, present invention effect in terms of objective quality is described further:

Table 1 is the present invention and the Y-PSNR (PeakSignaltoNoiseRatio, PSNR) of prior art method for video coding based on compressed sensing and the comparing result of scramble time.Realize the simulated conditions of this table: hardware environment: CPUAMDSempron3000+, 1.8GHz, 512MB internal memory；Software environment: WindowsXP, MatlabR2010b；Video block size: 8 × 8；Reference sequences: Carphone, News, Hall_monitor, Mother_daughter, Stefan；The frame number of cycle tests: the first frame of all sequences；Resolution: 176 × 144 (Carphone, News), 352 × 288 (Hall_monitor, Mother_daughter, Stefan)；Calculation matrix: gaussian random calculation matrix；Restructing algorithm: base tracing algorithm；Sparse base: the dictionary of adjacent reconstructing video frame structure.

As can be seen from Table 1, the present invention is compared with existing video coding technique based on compressed sensing, and the Y-PSNR PSNR of its decoding and reconstituting frame of video improves about 0.3～1.3dB.Owing to the encoding and decoding time increased that introduces of block sort method herein is less than the 5% of the whole system encoding and decoding time, there is no the complexity of increase system.Overall measurement number used by two kinds of method frame of video is essentially equal again, and therefore the present invention improves the distortion performance of whole system.

Table 1

Parameter in above-mentioned Fig. 3 and form demonstrates video compress perceptual coding system based on measurement field block sort and the method that the present invention proposes further, compared with method for video coding based on compressed sensing with prior art, improve quality and the distortion performance of whole video coding system of reconstructing video.

Claims (2)

1. a video compress perceptual coding system based on measurement field block sort, including splits' positions sampling module, the block sort module of measurement field, sample rate distribution module, hits distribution and double measurement module and reconstructed module；Input video block is divided into the sub-block of non-overlapping copies equal in magnitude by splits' positions sampling module, measures each sub-block with identical random measurement matrix and obtains measurement field signal；The block sort module of measurement field by extracting measurement field characteristic image module, measurement field is constituted with frequency domain relational model module and judging module, the measurement field signal of sub-block splits' positions sampling obtained obtains measurement field characteristic image as the row vector of matrix, measurement field utilizes cross covariance to weigh the dependency of measurement field signal and corresponding frequency-region signal to frequency domain relational model module, according to video signal coefficient characteristic distributions after discrete cosine transform, establish the linear approximate relationship of measurement field signal and the dependency of corresponding frequency-region signal；Judging module utilizes measurement field characteristic image to carry out the judgement of block classification with the correlative relationship model of measurement field Yu frequency-region signal for theoretical foundation, it is specially the Cross-covariance of computation and measurement characteristic of field image, and show that by correlative relationship model the variance of element of Cross-covariance is for weighing the feature of video block, size according to variance yields realizes the type judgement of video block, by judging module, the video block of input is judged to edge block, texture block or smooth block；Then system is for court verdict and the feature of dissimilar video block itself, the sample rate different to the distribution of different types of video block；Hits distribution and double measurement module determine measurement number and calculation matrix according to sample rate and video block sizes, utilize calculation matrix that video block re-starts measurement and obtain measurement field signal, finally measurement field signal are transferred to reconstructed module and are reconstructed.

2. a video compress perceptual coding method based on measurement field block sort, the system being suitable for is with claim 1, it is characterised in that: the method includes:

Step 1. is by input picture block or video block x_bIt is divided into the sub-block of non-overlapping copies equal in magnitude, it is assumed that be divided into four sub-blocks x_b1、x_b2、x_b3、x_b4, corresponding DCT domain signal is θ₁、θ₂、θ₃、θ₄, the gaussian random calculation matrix of each sub-block formed objects is measured respectively and obtains measurement field signal y₁, y₂, y₃, y₄, measurement field signal is the column vector of m × 1；

Step 2. using the measurement field signal of sub-block as matrix row vector build measurement field characteristic image y=[y₁, y₂, y₃, y₄]^T；

It is as follows with the step of the correlative relationship model of frequency-region signal that step 3. sets up measurement field signal:

3.1) the correlative relationship model of measurement field signal and frequency-region signal is set up:

$C_y=\frac{n}{m}{C}_{\mathrm{\θ}}$

Wherein, C_yFor the Cross-covariance of measurement field characteristic image, n is the pixel count of video block, and m is for measuring number, C_θFor the Cross-covariance in DCT domain；

3.2) according to the character of calculation matrix, the relational model of measurement field Cross-covariance and frequency-region signal is set up:

$C_y=\frac{1}{m}{\mathrm{\θ\θ}}^T$

Wherein, C_yFor the Cross-covariance of measurement field characteristic image, m is for measuring number, and θ is DCT domain sample matrix；

3.3) according to video signal in the characteristic distributions of DCT domain, the relational model of frequency domain covariance matrix and frequency-region signal is set up:

$C_{\mathrm{\θ}}=\frac{1}{n}{\mathrm{\θ\θ}}^T$

Wherein, C_θFor the Cross-covariance in DCT domain, n is the pixel count of video block, and θ is DCT domain sample matrix；

3.4) by step 3.2) and 3.3) relational model set up the correlative relationship model of measurement field signal and corresponding frequency-region signal:

$C_y=\frac{n}{m}{C}_{\mathrm{\θ}}$

Wherein, C_yFor the Cross-covariance of measurement field characteristic image, n is the pixel count of video block, and m is for measuring number, C_θFor the Cross-covariance in DCT domain；

Each element of the Cross-covariance of measurement field characteristic image is sought variance by step 4., the relational model of step 3 obtain:

$\mathrm{var}\left(C_y\right)={\left(\frac{n}{m}\right)}^2\·\mathrm{var}\left(C_{\mathrm{\θ}}\right)$

Wherein, variance, C are asked in var () expression_yFor the Cross-covariance of measurement field characteristic image, n is the pixel count of video block, and m is for measuring number, C_θFor the Cross-covariance in DCT domain；

The type of video block is made decisions by step 5. by following criterion, according to the variance size of the Cross-covariance element of measurement field characteristic image video block is judged to smooth block, edge block or texture block:

Wherein, if be " if " the meaning, var () represent ask variance, C_yFor the Cross-covariance of measurement field characteristic image, T₁, T₂For empirical value；

Step 6. carries out sample rate distribution to the video block of judgement, the distribution of sample rate meets edge block ＞ texture block ＞ smooth block, size according to sample rate and video block determines hits and calculation matrix, then utilizes this calculation matrix that current video block re-starts measurement and obtains measurement field signal y_b；

Step 7. is by measurement field signal y_bIt is transferred to reconstructed module, utilizes compressed sensing restructing algorithm that current video block is reconstructed.