CN104200498A - Real-time video super-resolution processing method integrated with cortex-A7 - Google Patents

Abstract

The invention discloses a real-time video super-resolution processing method integrated with cortex-A7. The method includes that (1), video sampling is performed, and a low-resolution video frame is obtained and input to a system on chip (SOC); (2) the low-resolution video frame is sequentially subjected to complexity processing, feature vector extracting and sample set training, feature vectors to be matched are obtained, and a sample set is established by high-resolution high-frequency components; (3) according to the feature vectors to be matched, the low-resolution video frame is subjected to super-resolution processing by means of an improved super-resolution algorithm based on cluster dictionary self-learning and feature sparse representation and combination with SOC coding and decoding technologies, and thereby, a high-resolution video frame flow is output. The real-time video super-resolution processing method integrated with cortex-A7 has the advantages of real time, low distortion rate and processing costs and high processing speed and qualities, and the method can be widely applied to the field of video image processing.

Description

Merge the real-time video super-resolution processing method of Cortex-A7

Technical field

The present invention relates to field of video image processing, especially merge the real-time video super-resolution processing method of Cortex-A7.

Background technology

At present, for most of imaging devices, its image resolution ratio of obtaining is also very low, and more exchange device needs a large amount of material resources and human input.Simultaneously in video image acquisition, owing to being subject to the impact of the many factors such as the distance of imaging device precision or equipment and target, the motion of target and noise, it obtains normally has noise, the video image that fuzzy and resolution is lower, is but difficult to obtain a width desired resolution image.Limited image resolution ratio can have influence on the performance of system, as low-resolution image can reduce the recognition performance of system.This often brings difficulty to work such as target identification, identity identification or criminal investigations, cannot meet actual demand.Therefore,, in the industry in the urgent need to studying a kind of new super-resolution technique, the method that the some frame low-resolution images under Same Scene can be processed by signal reverts to a vertical frame dimension image in different resolution, to reduce the cost of equipment.

The super-resolution algorithms of main flow comprises algorithm and the method based on rebuilding based on interpolation at present.Wherein, the algorithm based on interpolation, has lower algorithm complex, but it does not utilize the prior imformation of image, causes the image of recovery excessively level and smooth.And method based on rebuilding is limited to quantity and the wrong registration of low-resolution image, adaptability is poor.Utilize Gauss Markov Random Field Mixture model can improve this situation as the prior imformation of image.Yet when the limited amount of low-resolution image, the super-resolution algorithms based on Gauss Markov Random Field Mixture model is easily lost important detailed information, and distortion rate is higher.In addition, existing current super-resolution algorithms is still confined to process in single image, and the technology of processing live video stream is not yet ripe.If still continue to use existing super-resolution algorithms, process live video stream, can cause that its processing speed is slow, processing cost is higher and Disposal quality is lower.

Summary of the invention

In order to solve the problems of the technologies described above, the object of the invention is: provide a kind of in real time, distortion rate is lower, processing speed is very fast, processing cost is lower and quality is higher, the real-time video super-resolution processing method of fusion Cortex-A7.

The technical solution adopted for the present invention to solve the technical problems is:

The real-time video super-resolution processing method that merges Cortex-A7, comprising:

A, carry out video sampling, obtain low resolution video frame and be input in SOC SOC (system on a chip);

B, low resolution video frame is carried out successively to complexity processing, proper vector are extracted and sample set training, thereby obtain the proper vector that need to mate, described sample set adopts high-resolution high fdrequency component to build and forms;

C, the proper vector of mating as required, adopt improved super-resolution algorithms based on the self study of cluster dictionary and feature rarefaction representation and in conjunction with the encoding and decoding technique of SOC SOC (system on a chip), low resolution video frame is carried out to SUPERRESOLUTION PROCESSING FOR ACOUSTIC, thus output high-resolution video frame stream.

Further, described step C, it comprises:

C1, structure possess the improved super-resolution algorithms based on the self study of cluster dictionary and feature rarefaction representation of main feature;

The super-resolution algorithms of C2, the proper vector of mating as required and structure is carried out SUPERRESOLUTION PROCESSING FOR ACOUSTIC to low resolution video frame, thus output high-resolution video frame stream.

Further, described step C1, it comprises:

C11, set up one and cross complete data storehouse;

The rarefaction representation coefficient of the low-resolution video image block of C12, calculating input;

Sparse coding coefficient under C13, calculating low resolution dictionary and the sparse coding coefficient under high resolving power dictionary;

C14, according to the rarefaction representation coefficient of low-resolution video image block, sparse coding coefficient, high-definition picture storehouse and the low resolution dictionary crossed in complete data storehouse, reconstruct high resolution video image piece;

C15, adopt clustering algorithm and principal component analysis (PCA) to extract video image set of blocks, then adopt K-SVD algorithm to high-resolution and low-resolution image block set carry out joint training;

C16, according to the result of joint training, adopt orthogonal matching pursuit method to obtain possessing the improved super-resolution algorithms based on the self study of cluster dictionary and feature rarefaction representation of main feature.

Further, described step C13, it is specially:

Sparse coding coefficient under calculating low resolution dictionary and the sparse coding coefficient under high resolving power dictionary, the sparse coding coefficient under described low resolution dictionary with the sparse coding coefficient under high resolving power dictionary computing formula be respectively:

${\mathrm{\δ}}_L^{\mathrm{\ω}}=\arg \min {\left|\right|N_L^{\mathrm{\ω}}-K_L{\mathrm{\δ}}_L^{\mathrm{\ω}}\left|\right|}_2^{\mathrm{\ρ}}+\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}\left|\right|}_1,$

${\mathrm{\δ}}_H^{\mathrm{\ω}}=\arg \min {\left|\right|N_H^{\mathrm{\ω}}-K_H{\mathrm{\δ}}_H^{\mathrm{\ω}}\left|\right|}_2^{\mathrm{\ρ}}+\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_H^{\mathrm{\ω}}}\left|\right|}_1,$

Wherein, K _lfor low resolution dictionary function, represent the video flowing in low-resolution image, ρ is the given parameters of matrix norm, expression is desirable low resolution dictionary function bound term after deriving, and ε is the complicated regularization parameter after characteristic processing, and L is for replacing the norm of sparse coding;

K _hfor high resolving power dictionary function, represent the video flowing in high-definition picture, expression is desirable high resolving power dictionary function bound term after deriving, and H is for replacing the norm of sparse coding.

Further, described step C15, it is specially:

Adopt clustering algorithm and principal component analysis (PCA) to extract video image set of blocks, then adopt K-SVD algorithm to high-resolution and low-resolution image block set carry out joint training, thereby obtain the result data of joint training, the result data { K of described joint training _h, K _l, δ, ω } be:

Wherein, N is the column vector obtaining after high-resolution and low-resolution image block joins together to train, $N={[N_H\sec \left(N_H\right)/e^{\mathrm{i\ω}{w}_1},N_L\sec \left(N_L\right)/e^{\mathrm{i\ω}{w}_2}]}^T,$ $K={[K_H\sec \left(K_H\right)/e^{\mathrm{i\ω}{w}_1},K_L\sec \left(K_L\right)/e^{\mathrm{i\ω}{w}_2}]}^T$ For high low resolution associating dictionary, w ₁and w ₂be respectively by the dimension of the column vector of the high low-resolution video stream after training, δ=[δ ₁, δ ₂, δ ₃, δ ₄..., δ ₅] be code coefficient matrix, $\mathrm{\ω}=\left|\right|\begin{array}{cc}{\mathrm{\ω}}_1& {\mathrm{\ω}}_2\\ {\mathrm{\ω}}_3& {\mathrm{\ω}}_4\end{array}\left|\right|$ For desorption coefficient matrix.

Further, described step C2, it comprises:

C21, the proper vector that needs are mated are mated in dictionary database, and whether judgement coupling is successful, if so, performs step C23, otherwise, perform step C22;

C22, adopt improved K value iterative algorithm to carry out the self study of cluster dictionary to low resolution video frame, then perform step C23;

C23, according to bound term characteristic number coefficient, from cross complete data storehouse, finds out at a high speed and train dictionary library, and the video coding and decoding technology that merges SOC SOC (system on a chip) is optimized integrated processing to low resolution video frame, thereby export high-quality high-definition video stream.

Further, described step C22, it comprises:

C221, existing frame of video is carried out to training sample processing, thereby obtain elementary clustering function formula, the expression formula of described elementary clustering function is:

Wherein, I is dictionary type, and C is elementary cluster coefficients, for iterations, n is constant coefficient, for the reference data of training sample, for cluster coefficient of variation;

C222, from first pixel in every frame video image upper left corner, every a pixel, get a video image blocks, and got video image blocks adopted to the optimum solution of the elementary clustering function of LASSO Algorithm for Solving the optimum solution of described elementary clustering function expression formula be:

${\mathrm{\δ}}_{L1}^{\mathrm{\ω}}*\▿=\arg \min {\left|\right|N_L^{\mathrm{\ω}}-K_L{\mathrm{\δ}}_L^{\mathrm{\ω}}\left|\right|}_2^{\mathrm{\ρ}}+\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}\left|\right|}_1\underset{L}{\overset{\∈}{\mathrm{\Π}}}\underset{0\≤x\≤1}{\max }\tilde{X}{e}^{-x^2};$

Wherein, ▽ is LASSO operator, for the optimum integrated coefficient of super-resolution, expression formula be:

$\tilde{X}=\arg {\min }_{\mathrm{\α}}{\mathrm{\Σ}}_i{\left|\right|R^i(X_H-X_L)-\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}\left|\right|}_1\left|\right|}_6^8;$

C223, video image blocks is classified, obtain K cluster, then from each cluster learning, go out a sub-dictionary, thereby obtain optimum integrated coefficient under K sub-dictionary.

Further, described step C23, it comprises:

C231, the optimum solution to elementary clustering function carry out feature sparse coding average value constraint and process, thereby obtain adding the objective function after bound term characteristic number coefficient described objective function expression formula be:

$\begin{array}{c}{\mathrm{\δ}}_{L2}^{\mathrm{\ω}}=\arg \min {\left|\right|N_L^{\mathrm{\ω}}-K_L{\mathrm{\δ}}_L^{\mathrm{\ω}}\left|\right|}_2^{\mathrm{\ρ}}+\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}\left|\right|}_1\\ +\mathrm{\θ}{\left|\right|({\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}){\mathrm{\Σ}}_{n=1}^{\∞}({\mathrm{\δ}}_n\cos \frac{\mathrm{n\π\ω}}{{\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}}+\sin \frac{\mathrm{n\π\ω}}{{\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}})\left|\right|}_{\mathrm{\ϵ}}^2\end{array},$

Wherein, θ is feature constant, and $\mathrm{\θ}=\arg {\log }_5\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}+N_L^{\mathrm{\ω}}\left|\right|,$

$({\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}){\mathrm{\Σ}}_{n=1}^{\∞}({\mathrm{\δ}}_n\cos \frac{\mathrm{n\π\ω}}{{\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}}+\sin \frac{\mathrm{n\π\ω}}{{\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}})$ For representing with the characteristic quantity of proper vector distance;

The optimization average of C232, calculating cluster dictionary learning and sparse coding described optimization average computing formula be;

${{\mathrm{\Δ}}^{\mathrm{\ω}}}_1^T={\mathrm{\Σ}}_{(K_H,K_L,\mathrm{\δ},\mathrm{\ω})\∈N}{\mathrm{\δ}}_L^{\mathrm{\ω}}*\▿\{K_H,K_L,\mathrm{\δ},\mathrm{\ω}\},$

C233, to adding the objective function after bound term characteristic number coefficient with optimization average merge optimization process, thereby be optimized integrated objective function the integrated objective function of described optimization expression formula be:

$\begin{array}{c}{\mathrm{\δ}}_{L4}^{\mathrm{\ω}}=\arg \min {\left|\right|N_L^{\mathrm{\ω}}-K_L{\mathrm{\δ}}_L^{\mathrm{\ω}}\left|\right|}_2^{\mathrm{\ρ}}+\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}\left|\right|}_1\\ +\sqrt[T]{{\mathrm{\Σ}}_{(K_H,K_L,\mathrm{\δ},\mathrm{\ω})\∈N}{\mathrm{\δ}}_L^{\mathrm{\ω}}*\▿\{K_H,K_L,\mathrm{\δ},\mathrm{\ω}\}}\end{array};$

C234, according to optimizing integrated objective function, low resolution video frame is optimized to integrated processing, thereby generates the high-definition video stream line output of going forward side by side.

Further, described step C232, it comprises:

S1, employing zero-mean random variable are to adding the objective function after bound term characteristic number coefficient be out of shape, thus the objective function after being out of shape objective function after described distortion for:

${\mathrm{\δ}}_{L3}^{\mathrm{\ω}}=\arg \min {\left|\right|N_L^{\mathrm{\ω}}-K_L{\mathrm{\δ}}_L^{\mathrm{\ω}}\left|\right|}_2^{\mathrm{\ρ}}+\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}\left|\right|}_1+{\mathrm{\Δ}}^{\mathrm{\ω}},$

Wherein, Δ ^ωfor representing $\mathrm{\θ}{\left|\right|({\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}){\mathrm{\Σ}}_{n=1}^{\∞}({\mathrm{\δ}}_n\cos \frac{\mathrm{n\π\ω}}{{\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}}+\sin \frac{\mathrm{n\π\ω}}{{\mathrm{\δ}}_K^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}})\left|\right|}_{\mathrm{\ϵ}}^2;$

S2, according to the objective function after distortion, calculate the optimization average of cluster dictionary learning and sparse coding

The invention has the beneficial effects as follows: by cheap and good-quality SOC SOC (system on a chip), realize real-time video SUPERRESOLUTION PROCESSING FOR ACOUSTIC, processing cost is lower; Adopt the improved super-resolution algorithms based on the self study of cluster dictionary and feature rarefaction representation to carry out real-time video SUPERRESOLUTION PROCESSING FOR ACOUSTIC and in conjunction with the encoding and decoding technique of SOC SOC (system on a chip), solved the problem of the poor effect of processing live video stream, had advantages of in real time, distortion rate is lower, processing speed is very fast, processing cost is lower and quality is higher.

Accompanying drawing explanation

Below in conjunction with drawings and Examples, the invention will be further described.

Fig. 1 is the overall flow figure that the real-time video super-resolution processing method of Cortex-A7 is merged in the present invention;

Fig. 2 is the process flow diagram of step C of the present invention;

Fig. 3 is the process flow diagram of step C1 of the present invention;

Fig. 4 is the process flow diagram of step C2 of the present invention;

Fig. 5 is the process flow diagram of step C22 of the present invention;

Fig. 6 is the process flow diagram of step C23 of the present invention;

Fig. 7 is the process flow diagram of step C232 of the present invention

Fig. 8 is the hardware module structural drawing in the embodiment of the present invention one;

Fig. 9 is the algorithm flow chart of the embodiment of the present invention two.

Embodiment

With reference to Fig. 1, merge the real-time video super-resolution processing method of Cortex-A7, comprising:

A, carry out video sampling, obtain low resolution video frame and be input in SOC SOC (system on a chip);

B, low resolution video frame is carried out successively to complexity processing, proper vector are extracted and sample set training, thereby obtain the proper vector that need to mate, described sample set adopts high-resolution high fdrequency component to build and forms;

C, the proper vector of mating as required, adopt improved super-resolution algorithms based on the self study of cluster dictionary and feature rarefaction representation and in conjunction with the encoding and decoding technique of SOC SOC (system on a chip), low resolution video frame is carried out to SUPERRESOLUTION PROCESSING FOR ACOUSTIC, thus output high-resolution video frame stream.

Wherein, carry out complexity processing and proper vector and extract, for obtaining texture and the geometry feature of low-resolution image.A plurality of features that the object that carries out complexity processing is single image or the feature of multi-frame video image.

With reference to Fig. 2, be further used as preferred embodiment, described step C, it comprises:

C1, structure possess the improved super-resolution algorithms based on the self study of cluster dictionary and feature rarefaction representation of main feature;

The super-resolution algorithms of C2, the proper vector of mating as required and structure is carried out SUPERRESOLUTION PROCESSING FOR ACOUSTIC to low resolution video frame, thus output high-resolution video frame stream.

With reference to Fig. 3, be further used as preferred embodiment, described step C1, it comprises:

C11, set up one and cross complete data storehouse;

The rarefaction representation coefficient of the low-resolution video image block of C12, calculating input;

Sparse coding coefficient under C13, calculating low resolution dictionary and the sparse coding coefficient under high resolving power dictionary;

C14, according to the rarefaction representation coefficient of low-resolution video image block, sparse coding coefficient, high-definition picture storehouse and the low resolution dictionary crossed in complete data storehouse, reconstruct high resolution video image piece;

C15, adopt clustering algorithm and principal component analysis (PCA) to extract video image set of blocks, then adopt K-SVD algorithm to high-resolution and low-resolution image block set carry out joint training;

C16, according to the result of joint training, adopt orthogonal matching pursuit method to obtain possessing the improved super-resolution algorithms based on the self study of cluster dictionary and feature rarefaction representation of main feature.

Wherein, cross complete data storehouse, for storing the high-definition picture of training sample set, iteration and training dictionary etc.

K-SVD algorithm, be a kind of dictionary training algorithm of classics, according to error minimum principle, error term carried out to SVD decomposition, then select to make the decomposition item of error minimum as the dictionary atom upgrading and corresponding atom coefficient, and the solution being finally optimized through continuous iteration.

Be further used as preferred embodiment, described step C13, it is specially:

Sparse coding coefficient under calculating low resolution dictionary and the sparse coding coefficient under high resolving power dictionary, the sparse coding coefficient under described low resolution dictionary with the sparse coding coefficient under high resolving power dictionary computing formula be respectively:

${\mathrm{\δ}}_L^{\mathrm{\ω}}=\arg \min {\left|\right|N_L^{\mathrm{\ω}}-K_L{\mathrm{\δ}}_L^{\mathrm{\ω}}\left|\right|}_2^{\mathrm{\ρ}}+\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}\left|\right|}_1,$

${\mathrm{\δ}}_H^{\mathrm{\ω}}=\arg \min {\left|\right|N_H^{\mathrm{\ω}}-K_H{\mathrm{\δ}}_H^{\mathrm{\ω}}\left|\right|}_2^{\mathrm{\ρ}}+\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_H^{\mathrm{\ω}}}\left|\right|}_1,$

Wherein, K _lfor low resolution dictionary function, represent the video flowing in low-resolution image, ρ is the given parameters of matrix norm, expression is desirable low resolution dictionary function bound term after deriving, and ε is the complicated regularization parameter after characteristic processing, and L is for replacing the norm of sparse coding;

K _hfor high resolving power dictionary function, represent the video flowing in high-definition picture, expression is desirable high resolving power dictionary function bound term after deriving, and H is for replacing the norm of sparse coding.

Wherein, arg function is used for asking argument of complex number.ρ is the given parameters of matrix norm, according to the actual conditions of video flowing, sets.ε is the complicated regularization parameter after characteristic processing, is mainly used to balance K _lwith between there is ratio.

Wherein, L and H are all used to replace the norm of sparse coding, as the auxiliary function of derivation formula.

Be further used as preferred embodiment, described step C15, it is specially:

Adopt clustering algorithm and principal component analysis (PCA) to extract video image set of blocks, then adopt K-SVD algorithm to high-resolution and low-resolution image block set carry out joint training, thereby obtain the result data of joint training, the result data { K of described joint training _h, K _l, δ, ω } be:

Wherein, N is the column vector obtaining after high-resolution and low-resolution image block joins together to train,

$N={[N_H\sec \left(N_H\right)/e^{\mathrm{i\ω}{w}_1},N_L\sec \left(N_L\right)/e^{\mathrm{i\ω}{w}_2}]}^T,$

$K={[K_H\sec \left(K_H\right)/e^{\mathrm{i\ω}{w}_1},K_L\sec \left(K_L\right)/e^{\mathrm{i\ω}{w}_2}]}^T$ For high low resolution associating dictionary, w ₁and w ₂be respectively by the dimension of the column vector of the high low-resolution video stream after training, δ=[δ ₁, δ ₂, δ ₃, δ ₄..., δ ₅] be code coefficient matrix, $\mathrm{\ω}=\left|\right|\begin{array}{cc}{\mathrm{\ω}}_1& {\mathrm{\ω}}_2\\ {\mathrm{\ω}}_3& {\mathrm{\ω}}_4\end{array}\left|\right|$ For desorption coefficient matrix.

Wherein, sec function is secant trigonometric function.

With reference to Fig. 4, be further used as preferred embodiment, described step C2, it comprises:

C21, the proper vector that needs are mated are mated in dictionary database, and whether judgement coupling is successful, if so, performs step C23, otherwise, perform step C22;

C22, adopt improved K value iterative algorithm to carry out the self study of cluster dictionary to low resolution video frame, then perform step C23;

C23, according to bound term characteristic number coefficient, from cross complete data storehouse, finds out at a high speed and train dictionary library, and the video coding and decoding technology that merges SOC SOC (system on a chip) is optimized integrated processing to low resolution video frame, thereby export high-quality high-definition video stream.

With reference to Fig. 5, be further used as preferred embodiment, described step C22, it comprises:

C221, existing frame of video is carried out to training sample processing, thereby obtain elementary clustering function formula, the expression formula of described elementary clustering function is:

Wherein, I is dictionary type, and C is elementary cluster coefficients, for iterations, n is constant coefficient, for the reference data of training sample, for cluster coefficient of variation;

C222, from first pixel in every frame video image upper left corner, every a pixel, get a video image blocks, and got video image blocks adopted to the optimum solution of the elementary clustering function of LASSO Algorithm for Solving the optimum solution of described elementary clustering function expression formula be:

${\mathrm{\δ}}_{L1}^{\mathrm{\ω}}*\▿=\arg \min {\left|\right|N_L^{\mathrm{\ω}}-K_L{\mathrm{\δ}}_L^{\mathrm{\ω}}\left|\right|}_2^{\mathrm{\ρ}}+\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}\left|\right|}_1\underset{L}{\overset{\∈}{\mathrm{\Π}}}\underset{0\≤x\≤1}{\max }\tilde{X}{e}^{-x^2};$

Wherein, ▽ is LASSO operator, for the optimum integrated coefficient of super-resolution, expression formula be:

$\tilde{X}=\arg {\min }_{\mathrm{\α}}{\mathrm{\Σ}}_i{\left|\right|R^i(X_H-X_L)-\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}\left|\right|}_1\left|\right|}_6^8;$

C223, video image blocks is classified, obtain K cluster, then from each cluster learning, go out a sub-dictionary, thereby obtain optimum integrated coefficient under K sub-dictionary.

With reference to Fig. 6, be further used as preferred embodiment, described step C23, it comprises:

C231, the optimum solution to elementary clustering function carry out feature sparse coding average value constraint and process, thereby obtain adding the objective function after bound term characteristic number coefficient described objective function expression formula be:

$\begin{array}{c}{\mathrm{\δ}}_{L2}^{\mathrm{\ω}}=\arg \min {\left|\right|N_L^{\mathrm{\ω}}-K_L{\mathrm{\δ}}_L^{\mathrm{\ω}}\left|\right|}_2^{\mathrm{\ρ}}+\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}\left|\right|}_1\\ +\mathrm{\θ}{\left|\right|({\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}){\mathrm{\Σ}}_{n=1}^{\∞}({\mathrm{\δ}}_n\cos \frac{\mathrm{n\π\ω}}{{\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}}+\sin \frac{\mathrm{n\π\ω}}{{\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}})\left|\right|}_{\mathrm{\ϵ}}^2\end{array},$

Wherein, θ is feature constant, and $\mathrm{\θ}=\arg {\log }_5\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}+N_L^{\mathrm{\ω}}\left|\right|,$

$({\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}){\mathrm{\Σ}}_{n=1}^{\∞}({\mathrm{\δ}}_n\cos \frac{\mathrm{n\π\ω}}{{\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}}+\sin \frac{\mathrm{n\π\ω}}{{\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}})$ For representing with the characteristic quantity of proper vector distance;

The optimization average of C232, calculating cluster dictionary learning and sparse coding described optimization average computing formula be;

${{\mathrm{\Δ}}^{\mathrm{\ω}}}_1^T={\mathrm{\Σ}}_{(K_H,K_L,\mathrm{\δ},\mathrm{\ω})\∈N}{\mathrm{\δ}}_L^{\mathrm{\ω}}*\▿\{K_H,K_L,\mathrm{\δ},\mathrm{\ω}\},$

C233, to adding the objective function after bound term characteristic number coefficient with optimization average merge optimization process, thereby be optimized integrated objective function the integrated objective function of described optimization expression formula be:

$\begin{array}{c}{\mathrm{\δ}}_{L4}^{\mathrm{\ω}}=\arg \min {\left|\right|N_L^{\mathrm{\ω}}-K_L{\mathrm{\δ}}_L^{\mathrm{\ω}}\left|\right|}_2^{\mathrm{\ρ}}+\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}\left|\right|}_1\\ +\sqrt[T]{{\mathrm{\Σ}}_{(K_H,K_L,\mathrm{\δ},\mathrm{\ω})\∈N}{\mathrm{\δ}}_L^{\mathrm{\ω}}*\▿\{K_H,K_L,\mathrm{\δ},\mathrm{\ω}\}}\end{array};$

C234, according to optimizing integrated objective function, low resolution video frame is optimized to integrated processing, thereby generates the high-definition video stream line output of going forward side by side.

With reference to Fig. 7, be further used as preferred embodiment, described step C232, it comprises:

S1, employing zero-mean random variable are to adding the objective function after bound term characteristic number coefficient be out of shape, thus the objective function after being out of shape objective function after described distortion for:

${\mathrm{\δ}}_{L3}^{\mathrm{\ω}}=\arg \min {\left|\right|N_L^{\mathrm{\ω}}-K_L{\mathrm{\δ}}_L^{\mathrm{\ω}}\left|\right|}_2^{\mathrm{\ρ}}+\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}\left|\right|}_1+{\mathrm{\Δ}}^{\mathrm{\ω}},$

Wherein, Δ ^ωfor representing $\mathrm{\θ}{\left|\right|({\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}){\mathrm{\Σ}}_{n=1}^{\∞}({\mathrm{\δ}}_n\cos \frac{\mathrm{n\π\ω}}{{\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}}+\sin \frac{\mathrm{n\π\ω}}{{\mathrm{\δ}}_K^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}})\left|\right|}_{\mathrm{\ϵ}}^2;$

S2, according to the objective function after distortion, calculate the optimization average of cluster dictionary learning and sparse coding

Below in conjunction with Figure of description and specific embodiment, the present invention is described in further detail.

Embodiment mono-

The present invention, for saving hardware cost and human cost, realizes the function of real-time video super-resolution, to meet the requirement to high-resolution imaging of enterprise and associated mechanisms by cheap and good-quality SOC SOC (system on a chip).

SOC SOC (system on a chip) of the present invention is comprised of hardware module and software module two parts.Wherein, hardware components chief component comprises: the high energy efficiency process chip of a. based on ARMv7-A framework; B.DDR3 processor and external unit thereof (camera, mouse, keyboard, display, power supply); C.SOC chip external interface (comprising high speed SDRAM data-interface, video data input interface, video data output interface, interrupt interface, DDR3 data-interface).Whole hardware module structure as shown in Figure 8.

In Fig. 8, 1 is the SOC (system on a chip) main control processor based on SOC, 2 is the data buffer storage chip synchronous DRAM of system, 3 is Double Data Rate synchronous DRAM, 4 is to be the data cache interface between SOC main control processor and SDRAM, 5 is to be the data cache interface between SOC main control processor and Double Data Rate synchronous DRAM, 6 is mouse data Data Input Interface, 7 is keyboard data control inputs interface, the 8th, be the steering order IO interface of whole system, the 9th, the data exchange interface of hard disc data and SOC (system on a chip), 10 is vedio data stream input interface, 11 display systems that carry for SOC SOC (system on a chip) are for the display interface of display screen, 12 for connecting the power lead of terminal display screen and the power lead that connects SOC master control development board, 13 is SOC system, the power supply of terminal display screen and other peripheral hardware, 14 is the terminal display screen of whole system, 15 provide the equipment of video acquisition data (as common camera for whole system, video camera, industry camera and industrial camera), 16 provide the hard disk that stores video data for whole SOC development system, 17 is data, services switching processor, 18 central processing units for whole SOC system and whole system mode of operation provide the equipment (as keyboard) of steering order input, 19 central processing units for whole SOC system and whole system mode of operation provide the external unit (as mouse) of steering order input.Wherein, for whole system provides the equipment of video acquisition data, the form of its collection is AVI or MPEG, video input interface receives the data stream from SOC chip, and data cache interface connects high-speed synchronous dynamic RAM, steering order equipment input interface receives from external unit mouse-keyboard and other outer steering order of asking hardware device, the display system that SOC SOC (system on a chip) carries supplies the display interface of display screen show data and control in real time sequential for terminal display screen provides, and whole system is in steady operation pattern.

Software module comprises this four major part of Ubuntu embedded system, OpenCV function library, super-resolution algorithms module and compiler.

During normal work, software systems, can independent operating SOC SOC (system on a chip) under the support of SOC hardware system.Wherein, Ubuntu embedded system, is a complete development system, can provide a stable and high performance development environment for whole system.OpenCV function library, a medium of supporting hardware system and software systems, it can, by having called and algorithm through manually training, carry out real-time and perfect function to development system and realize.Super-resolution algorithms module, has merged the core algorithm technology of the real-time video super-resolution of Cortex-A7, in the situation that hardware environment and software environment are all put up, by improved algorithmic technique, just can carry out complete realization to the present invention.Compiler, for program in machine code is write and compiled, is equivalent to a medium, can carry out effect realization to the project that completes and design.

The process that software module of the present invention realizes comprises: (1) Ubuntu embedded system is called the process of OpenCV function library; (2) super-resolution algorithms module is called the process of OpenCV function library; (3) compiler compiles project (algorithm having completed can be written as the form of project), and form can running program; (4) the complete process that realizes application program of instrument can running program carrying by Ubuntu system; (5) Ubuntu embedded system is installed the process of OpenCV function library or the process of recurrence; (6) algorithm after training compiles into the standby process in OpenCV storehouse; (7) compiler is to algorithm or project is debugged repeatedly or the experimentation of effect comparison; (8) for Ubuntu embedded system, upgrade or call the process of compiler.

Embodiment bis-

The real-time video super-resolution processing method of Cortex-A7 is merged in the present invention, traditional super-resolution treatment technology, conventional video coding and decoding technology and traditional image coding technique are combined closely, study and propose a kind of new improvement algorithm based on hardware and software platform.In the improvement algorithm coding algorithm of real-time video super-resolution that merges Cortex-A7, guarantee that it can be in the smooth and easy operation of hardware platform SOC SOC (system on a chip) and best results, key is video coding and decoding technology and processes the objective quality of video image algorithmic technique.Because only under identical decoded picture Y-PSNR (Peak signal to noise ratio, PSNR) prerequisite, bit rate has reduced, its code decode algorithm just has cogency.And affect the Theoretical Mass that one of decoding video images objective quality key factor is the video image after SUPERRESOLUTION PROCESSING FOR ACOUSTIC.Algorithm after the present invention improves carries out to correlation parameter and video sampling image sequence the super-resolution algorithms experiment repeating for several times in decoding end, through under the experiment condition of various complex environments, obtains final decoding video images.The present invention is merged to the improvement algorithm fusion of real-time video super-resolution of Cortex-A7 in video encoding standard MPEG-2, test result shows, identical PSNR relatively under, its bit rate decrease to some degree, the algorithm after improvement possesses very strong practicality.In addition, owing to having used the brand-new coding and decoding video that passes through SOC SOC (system on a chip) in the algorithm after improving and in conjunction with the super-resolution algorithms technology of the self study of improved cluster dictionary and feature rarefaction representation, the super-resolution algorithms after this improvement can be fused in the encryption algorithm in standard card cage H.264/AVC.Algorithm after this improvement carries out down-sampling coding at coding side to original video, at the super-resolution algorithms technical finesse B of the improved cluster dictionary learning of decoding end and feature rarefaction representation frame (bi-directional predicted interpolation coding frame) and P frame (forward-predictive-coded frames).

Embodiment tri-

OpenCV function library has the multiple interpolating function carrying, and also carry the aggregation function based on variational algorithm, but experiment effect is unsatisfactory simultaneously.The super-resolution technique of the dictionary of main flow study in the recent period and rarefaction representation does not have in OpenCV function library.Therefore, under the pressure of the demand of market economy and the needs of technical development, the present invention proposes an improved new algorithm, is embedded into OpenCV function library, to reach the demand of real-time video SUPERRESOLUTION PROCESSING FOR ACOUSTIC transplantable, real-time, SOC SOC (system on a chip).

The present embodiment is elaborated to the brand-new coding and decoding video that passes through SOC SOC (system on a chip) of the present invention and in conjunction with the super-resolution algorithms of improved cluster dictionary learning and feature rarefaction representation.

The super-resolution algorithms of improved cluster dictionary learning and feature rarefaction representation is a kind of new practical algorithm proposing after improving on the basis of super-resolution rebuilding technology based on rarefaction representation and the super-resolution based on dictionary learning.Suitable complete data storehouse excessively of this algorithm model, then calculate the rarefaction representation coefficient of the low-resolution video image block of input, then calculate high resolving power and low resolution dictionary, the process of training is by these coefficients, high-definition picture storehouse and low resolution dictionary reconstruct high resolution video image piece, utilize clustering algorithm and principal component analysis (PCA) (PCA) to extract the part in video image set of blocks, and adopt K-SVD method to height, low-resolution image set of blocks carries out joint training, in super-resolution image processing procedure, take the method for orthogonal matching pursuit OMP to obtain the improved cluster dictionary learning of self study and the super-resolution algorithms of feature rarefaction representation, construct the super-resolution algorithms that possesses the improved cluster dictionary learning of main feature and feature rarefaction representation.The method has guaranteed that high-resolution and low-resolution video image blocks represents the consistance of coefficient completely, has also reduced the complexity of rebuilding simultaneously, and the adaptivity that is conducive to improve cluster dictionary learning and feature rarefaction representation has shortened the time of training simultaneously.Algorithm of the present invention has good performance, and the result after SUPERRESOLUTION PROCESSING FOR ACOUSTIC has higher PSNR and MSSIM (average structure similarity).

In the present invention, the sparse coding coefficient table calculating after deriving under low resolution dictionary is shown:

${\mathrm{\δ}}_L^{\mathrm{\ω}}=\arg \min {\left|\right|N_L^{\mathrm{\ω}}-K_L{\mathrm{\δ}}_L^{\mathrm{\ω}}\left|\right|}_2^{\mathrm{\ρ}}+\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}\left|\right|}_1---\left(1\right)$

Through calculating, try to achieve low-resolution video flow data at dictionary K _lunder sparse coding coefficient can be expressed as with that the in the situation that of video image super-resolution, because high-definition video stream and low-resolution video stream possess identical rarefaction representation coefficient, therefore high resolution video image cluster dictionary can be quoted the derivation formula of low resolution, that is:

${\mathrm{\δ}}_H^{\mathrm{\ω}}=\arg \min {\left|\right|N_H^{\mathrm{\ω}}-K_H{\mathrm{\δ}}_H^{\mathrm{\ω}}\left|\right|}_2^{\mathrm{\ρ}}+\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_H^{\mathrm{\ω}}}\left|\right|}_1---\left(2\right)$

In order to allow consistance in derivation formula and that possess height, so that do more accurately and calculate for later work, next the present invention will carry out joint training to derivation formula on the basis in formula (1) and (2), and training result is as follows:

In the present invention, after above coefficient formula is derived successfully, next formally enter the core algorithm process of the super-resolution algorithms of improved cluster dictionary learning and feature rarefaction representation.

Improved cluster dictionary learning, by setting up digital model, is processed existing video, carries out the sample training of thousands of times, obtains as shown in the formula the elementary cluster coefficients formula shown in (4):

In formula (4), primary training is to learn out the sub-dictionary IH of low resolution cluster and IL, then gathers the frame of video of some, and these frame of video are carried out to clustering learning processing, and the sample set obtaining is as shown in the formula shown in (5) and formula (6):

${\ℵ}_H=\{{\ℵ}_H^1,{\ℵ}_H^2,{\ℵ}_H^3,{\ℵ}_H^4,{\ℵ}_H^5,......{\ℵ}_H^i\}---\left(5\right)$

${\ℵ}_L=\{{\ℵ}_L^1,{\ℵ}_L^2,{\ℵ}_L^3,{\ℵ}_L^4,{\ℵ}_L^5,......{\ℵ}_L^i\}---\left(6\right)$

In formula, for high-resolution sample set, sample set for low resolution; For I, from first pixel in the upper left corner of every frame video image, every a pixel, get a video image blocks.

Then, use LASSO Algorithm for Solving optimal problem, obtain following formula (7):

${\mathrm{\δ}}_{L1}^{\mathrm{\ω}}*\▿=\arg \min {\left|\right|N_L^{\mathrm{\ω}}-K_L{\mathrm{\δ}}_L^{\mathrm{\ω}}\left|\right|}_2^{\mathrm{\ρ}}+\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}\left|\right|}_1\underset{L}{\overset{\∈}{\mathrm{\Π}}}\underset{0\≤x\≤1}{\max }\tilde{X}{e}^{-x^2}---\left(7\right)$

Wherein, for the optimum integrated coefficient of super-resolution.This coefficient value as follows:

$\tilde{X}=\arg {\min }_{\mathrm{\α}}{\mathrm{\Σ}}_i{\left|\right|R^i(X_H-X_L)-\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}\left|\right|}_1\left|\right|}_6^8---\left(8\right)$

In order to obtain optimization coefficient k in a situation sub-dictionary, the present invention classifies video image blocks, obtains K cluster, then from each cluster learning, goes out a sub-dictionary.Obviously, K cluster can represent K different tactic pattern, thereby can obtain the cluster resolution function in perception meaning.

The present invention has adopted improved K value cluster dictionary learning to solve noise problem, has introduced the rear super-resolution algorithms of naturally wanting introduced feature rarefaction representation of K value.By feature sparse table coding average value constraint, improve the quality of algorithm.Wherein, formula (1) has guaranteed all the time be infinitely close to but cannot guarantee can unconditionally be infinitely close to condition and practical application by experiment, can prove that both exist feature sparse coding noise before

Therefore, need to reduce feature sparse coding noise, namely need using feature sparse coding noise as a new bound term.Add the objective function system after bound term to become:

$\begin{array}{c}{\mathrm{\δ}}_{L2}^{\mathrm{\ω}}=\arg \min {\left|\right|N_L^{\mathrm{\ω}}-K_L{\mathrm{\δ}}_L^{\mathrm{\ω}}\left|\right|}_2^{\mathrm{\ρ}}+\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}\left|\right|}_1\\ +\mathrm{\θ}{\left|\right|({\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}){\mathrm{\Σ}}_{n=1}^{\∞}({\mathrm{\δ}}_n\cos \frac{\mathrm{n\π\ω}}{{\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}}+\sin \frac{\mathrm{n\π\ω}}{{\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}})\left|\right|}_{\mathrm{\ϵ}}^2\end{array}---\left(10\right)$

Next adopt zero-mean random variable to be out of shape formula (10), the result of distortion is:

${\mathrm{\δ}}_{L3}^{\mathrm{\ω}}=\arg \min {\left|\right|N_L^{\mathrm{\ω}}-K_L{\mathrm{\δ}}_L^{\mathrm{\ω}}\left|\right|}_2^{\mathrm{\ρ}}+\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}\left|\right|}_1+{\mathrm{\Δ}}^{\mathrm{\ω}}---\left(11\right)$

The present invention adopts a kind of K value method of new weighted iteration to represent distance feature, and weight distance feature more hour is more obvious; Distance feature when weights are larger is fuzzyyer.Based on these distance feature, the present invention proposes a new formula (12) and calculates acquisition cluster dictionary learning and sparse coding optimization average:

${{\mathrm{\Δ}}^{\mathrm{\ω}}}_1^T={\mathrm{\Σ}}_{(K_H,K_L,\mathrm{\δ},\mathrm{\ω})\∈N}{\mathrm{\δ}}_L^{\mathrm{\ω}}*\▿\{K_H,K_L,\mathrm{\δ},\mathrm{\ω}\}---\left(12\right)$

After accurate calculating, the present invention can obtain the super-resolution complexity coefficient of high-precision feature rarefaction representation.

So far, the super-resolution algorithms of improved cluster dictionary learning and feature rarefaction representation has completed most of work, but consider homoorganicity problem, for algorithm stability reliability is further improved, the present invention obtains after formula (10) (12) is merged to optimization process:

$\begin{array}{c}{\mathrm{\δ}}_{L4}^{\mathrm{\ω}}=\arg \min {\left|\right|N_L^{\mathrm{\ω}}-K_L{\mathrm{\δ}}_L^{\mathrm{\ω}}\left|\right|}_2^{\mathrm{\ρ}}+\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}\left|\right|}_1\\ +\sqrt[T]{{\mathrm{\Σ}}_{(K_H,K_L,\mathrm{\δ},\mathrm{\ω})\∈N}{\mathrm{\δ}}_L^{\mathrm{\ω}}*\▿\{K_H,K_L,\mathrm{\δ},\mathrm{\ω}\}}\end{array}---\left(14\right).$

The process flow diagram of whole system of the present invention and algorithm as shown in Figure 9.

Embodiment tri-

The PSNR of algorithm of the present invention and each traditional algorithm (dB) and SSIM value are carried out to a contrast, as shown in table 1 below.The first behavior PSNR of each sample (Test1～Test5), the second behavior SSIM.

Table 1

As can be seen from Table 1, algorithm of the present invention can carry out super-resolution Video processing to video sequence image very in high quality, thus the high Video processing super-resolution image of quality obtaining.

SOC SOC (system on a chip) has been merged in the present invention, first by complexity, process with proper vector and extract texture and the geometry feature of obtaining low-resolution image, then adopt training set to carry out sample training, the last priori of having utilized well again image according to the result employing of training, can effectively prevent that the image recovering is excessively level and smooth; Comprised the process of carrying out complexity processing, and adopt the sample set of high-resolution high fdrequency component formation and the resolution super-resolution algorithms of feature rarefaction representation to process, the quantity of low-resolution image and the impact that wrong registration brings have been reduced, adaptability is better, and be difficult for losing important detailed information, distortion rate is lower.Actual experimental results shows, adopts method of the present invention to process the rear video streaming image quality obtaining higher, and result is more satisfactory.

More than that better enforcement of the present invention is illustrated, but the invention is not limited to described embodiment, those of ordinary skill in the art also can make all equivalent variations or replacement under the prerequisite without prejudice to spirit of the present invention, and the distortion that these are equal to or replacement are all included in the application's claim limited range.

Claims (9)

1. merge the real-time video super-resolution processing method of Cortex-A7, it is characterized in that: comprising:

A, carry out video sampling, obtain low resolution video frame and be input in SOC SOC (system on a chip);

B, low resolution video frame is carried out successively to complexity processing, proper vector are extracted and sample set training, thereby obtain the proper vector that need to mate, described sample set adopts high-resolution high fdrequency component to build and forms;

C, the proper vector of mating as required, adopt improved super-resolution algorithms based on the self study of cluster dictionary and feature rarefaction representation and in conjunction with the encoding and decoding technique of SOC SOC (system on a chip), low resolution video frame is carried out to SUPERRESOLUTION PROCESSING FOR ACOUSTIC, thus output high-resolution video frame stream.

2. the real-time video super-resolution processing method of fusion according to claim 1 Cortex-A7, is characterized in that: described step C, and it comprises:

C1, structure possess the improved super-resolution algorithms based on the self study of cluster dictionary and feature rarefaction representation of main feature;

The super-resolution algorithms of C2, the proper vector of mating as required and structure is carried out SUPERRESOLUTION PROCESSING FOR ACOUSTIC to low resolution video frame, thus output high-resolution video frame stream.

3. the real-time video super-resolution processing method of fusion according to claim 2 Cortex-A7, is characterized in that: described step C1, and it comprises:

C11, set up one and cross complete data storehouse;

The rarefaction representation coefficient of the low-resolution video image block of C12, calculating input;

Sparse coding coefficient under C13, calculating low resolution dictionary and the sparse coding coefficient under high resolving power dictionary;

C14, according to the rarefaction representation coefficient of low-resolution video image block, sparse coding coefficient, high-definition picture storehouse and the low resolution dictionary crossed in complete data storehouse, reconstruct high resolution video image piece;

C15, adopt clustering algorithm and principal component analysis (PCA) to extract video image set of blocks, then adopt K-SVD algorithm to high-resolution and low-resolution image block set carry out joint training;

C16, according to the result of joint training, adopt orthogonal matching pursuit method to obtain possessing the improved super-resolution algorithms based on the self study of cluster dictionary and feature rarefaction representation of main feature.

4. the real-time video super-resolution processing method of fusion according to claim 3 Cortex-A7, is characterized in that: described step C13, and it is specially:

Sparse coding coefficient under calculating low resolution dictionary and the sparse coding coefficient under high resolving power dictionary, the sparse coding coefficient under described low resolution dictionary with the sparse coding coefficient under high resolving power dictionary computing formula be respectively:

${\mathrm{\δ}}_L^{\mathrm{\ω}}=\arg \min {\left|\right|N_L^{\mathrm{\ω}}-K_L{\mathrm{\δ}}_L^{\mathrm{\ω}}\left|\right|}_2^{\mathrm{\ρ}}+\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}\left|\right|}_1,$

${\mathrm{\δ}}_H^{\mathrm{\ω}}=\arg \min {\left|\right|N_H^{\mathrm{\ω}}-K_H{\mathrm{\δ}}_H^{\mathrm{\ω}}\left|\right|}_2^{\mathrm{\ρ}}+\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_H^{\mathrm{\ω}}}\left|\right|}_1,$

Wherein, K _lfor low resolution dictionary function, represent the video flowing in low-resolution image, ρ is the given parameters of matrix norm, expression is desirable low resolution dictionary function bound term after deriving, and ε is the complicated regularization parameter after characteristic processing, and L is for replacing the norm of sparse coding;

K _hfor high resolving power dictionary function, represent the video flowing in high-definition picture, expression is desirable high resolving power dictionary function bound term after deriving, and H is for replacing the norm of sparse coding.

5. the real-time video super-resolution processing method of fusion according to claim 4 Cortex-A7, is characterized in that: described step C15, and it is specially:

Adopt clustering algorithm and principal component analysis (PCA) to extract video image set of blocks, then adopt K-SVD algorithm to high-resolution and low-resolution image block set carry out joint training, thereby obtain the result data of joint training, the result data { K of described joint training _h, K _l, δ, ω } be:

Wherein, N is the column vector obtaining after high-resolution and low-resolution image block joins together to train, $N={[N_H\sec \left(N_H\right)/e^{\mathrm{i\ω}{w}_1},N_L\sec \left(N_L\right)/e^{\mathrm{i\ω}{w}_2}]}^T,$ $K={[K_H\sec \left(K_H\right)/e^{\mathrm{i\ω}{w}_1},K_L\sec \left(K_L\right)/e^{\mathrm{i\ω}{w}_2}]}^T$ For high low resolution associating dictionary, w ₁and w ₂be respectively by the dimension of the column vector of the high low-resolution video stream after training, δ=[δ ₁, δ ₂, δ ₃, δ ₄..., δ ₅] be code coefficient matrix, $\mathrm{\ω}=\left|\right|\begin{array}{cc}{\mathrm{\ω}}_1& {\mathrm{\ω}}_2\\ {\mathrm{\ω}}_3& {\mathrm{\ω}}_4\end{array}\left|\right|$ For desorption coefficient matrix.

6. the real-time video super-resolution processing method of fusion according to claim 5 Cortex-A7, is characterized in that: described step C2, and it comprises:

C21, the proper vector that needs are mated are mated in dictionary database, and whether judgement coupling is successful, if so, performs step C23, otherwise, perform step C22;

C22, adopt improved K value iterative algorithm to carry out the self study of cluster dictionary to low resolution video frame, then perform step C23;

C23, according to bound term characteristic number coefficient, from cross complete data storehouse, finds out at a high speed and train dictionary library, and the video coding and decoding technology that merges SOC SOC (system on a chip) is optimized integrated processing to low resolution video frame, thereby export high-quality high-definition video stream.

7. the real-time video super-resolution processing method of fusion according to claim 6 Cortex-A7, is characterized in that: described step C22, and it comprises:

C221, existing frame of video is carried out to training sample processing, thereby obtain elementary clustering function formula, the expression formula of described elementary clustering function is:

Wherein, I is dictionary type, and C is elementary cluster coefficients, for iterations, n is constant coefficient, for the reference data of training sample, for cluster coefficient of variation;

C222, from first pixel in every frame video image upper left corner, every a pixel, get a video image blocks, and got video image blocks adopted to the optimum solution of the elementary clustering function of LASSO Algorithm for Solving the optimum solution of described elementary clustering function expression formula be:

${\mathrm{\δ}}_{L1}^{\mathrm{\ω}}*\▿=\arg \min {\left|\right|N_L^{\mathrm{\ω}}-K_L{\mathrm{\δ}}_L^{\mathrm{\ω}}\left|\right|}_2^{\mathrm{\ρ}}+\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}\left|\right|}_1\underset{L}{\overset{\∈}{\mathrm{\Π}}}\underset{0\≤x\≤1}{\max }\tilde{X}{e}^{-x^2};$

Wherein, ▽ is LASSO operator, for the optimum integrated coefficient of super-resolution, expression formula be:

$\tilde{X}=\arg {\min }_{\mathrm{\α}}{\mathrm{\Σ}}_i{\left|\right|R^i(X_H-X_L)-\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}\left|\right|}_1\left|\right|}_6^8;$

C223, video image blocks is classified, obtain K cluster, then from each cluster learning, go out a sub-dictionary, thereby obtain optimum integrated coefficient under K sub-dictionary.

8. the real-time video super-resolution processing method of fusion according to claim 7 Cortex-A7, is characterized in that: described step C23, and it comprises:

C231, the optimum solution to elementary clustering function carry out feature sparse coding average value constraint and process, thereby obtain adding the objective function after bound term characteristic number coefficient described objective function expression formula be:

$\begin{array}{c}{\mathrm{\δ}}_{L2}^{\mathrm{\ω}}=\arg \min {\left|\right|N_L^{\mathrm{\ω}}-K_L{\mathrm{\δ}}_L^{\mathrm{\ω}}\left|\right|}_2^{\mathrm{\ρ}}+\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}\left|\right|}_1\\ +\mathrm{\θ}{\left|\right|({\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}){\mathrm{\Σ}}_{n=1}^{\∞}({\mathrm{\δ}}_n\cos \frac{\mathrm{n\π\ω}}{{\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}}+\sin \frac{\mathrm{n\π\ω}}{{\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}})\left|\right|}_{\mathrm{\ϵ}}^2\end{array},$

Wherein, θ is feature constant, and $\mathrm{\θ}=\arg {\log }_5\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}+N_L^{\mathrm{\ω}}\left|\right|,$

$({\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}){\mathrm{\Σ}}_{n=1}^{\∞}({\mathrm{\δ}}_n\cos \frac{\mathrm{n\π\ω}}{{\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}}+\sin \frac{\mathrm{n\π\ω}}{{\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}})$ For representing with the characteristic quantity of proper vector distance;

The optimization average of C232, calculating cluster dictionary learning and sparse coding described optimization average computing formula be;

${{\mathrm{\Δ}}^{\mathrm{\ω}}}_1^T={\mathrm{\Σ}}_{(K_H,K_L,\mathrm{\δ},\mathrm{\ω})\∈N}{\mathrm{\δ}}_L^{\mathrm{\ω}}*\▿\{K_H,K_L,\mathrm{\δ},\mathrm{\ω}\},$

C233, to adding the objective function after bound term characteristic number coefficient with optimization average merge optimization process, thereby be optimized integrated objective function the integrated objective function of described optimization expression formula be:

$\begin{array}{c}{\mathrm{\δ}}_{L4}^{\mathrm{\ω}}=\arg \min {\left|\right|N_L^{\mathrm{\ω}}-K_L{\mathrm{\δ}}_L^{\mathrm{\ω}}\left|\right|}_2^{\mathrm{\ρ}}+\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}\left|\right|}_1\\ +\sqrt[T]{{\mathrm{\Σ}}_{(K_H,K_L,\mathrm{\δ},\mathrm{\ω})\∈N}{\mathrm{\δ}}_L^{\mathrm{\ω}}*\▿\{K_H,K_L,\mathrm{\δ},\mathrm{\ω}\}}\end{array};$

C234, according to optimizing integrated objective function, low resolution video frame is optimized to integrated processing, thereby generates the high-definition video stream line output of going forward side by side.

9. the real-time video super-resolution processing method of fusion according to claim 8 Cortex-A7, is characterized in that: described step C232, and it comprises:

S1, employing zero-mean random variable are to adding the objective function after bound term characteristic number coefficient be out of shape, thus the objective function after being out of shape objective function after described distortion for:

${\mathrm{\δ}}_{L3}^{\mathrm{\ω}}=\arg \min {\left|\right|N_L^{\mathrm{\ω}}-K_L{\mathrm{\δ}}_L^{\mathrm{\ω}}\left|\right|}_2^{\mathrm{\ρ}}+\frac{1}{15}\mathrm{\ϵ}{\left|\right|\sqrt[\mathrm{\ω}]{{\mathrm{\δ}}_L^{\mathrm{\ω}}}\left|\right|}_1+{\mathrm{\Δ}}^{\mathrm{\ω}},$

Wherein, Δ ^ωfor representing $\mathrm{\θ}{\left|\right|({\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}){\mathrm{\Σ}}_{n=1}^{\∞}({\mathrm{\δ}}_n\cos \frac{\mathrm{n\π\ω}}{{\mathrm{\δ}}_H^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}}+\sin \frac{\mathrm{n\π\ω}}{{\mathrm{\δ}}_K^{\mathrm{\ω}}-{\mathrm{\δ}}_L^{\mathrm{\ω}}})\left|\right|}_{\mathrm{\ϵ}}^2;$

S2, according to the objective function after distortion, calculate the optimization average of cluster dictionary learning and sparse coding