CN104616011B - MRF (Multi-Reference Frame) license plate denoising algorithm based on combined apriorism of gradient information and block area - Google Patents

Abstract

The invention belongs to the field of image information processing, and particularly relates to an MRF (Multi-Reference Frame) license plate denoising algorithm based on combined apriorism of gradient information and a block area. The method comprises the specific steps: S1, building license plate image apriorism information including the gradient apriorism and block area apriorism by using a number gradient information of a license plate; S2, carrying out noise estimation on a video image sequence, and determining a density function of noise distribution probability; S3, building a two-value markov random field model of the license plate image, combining the markov random field model with the license plate image apriorism information, and building an optimization problem model and solving. Compared with a traditional license plate denoising method, the SNR (Signal to Noise Ratio) of the license plate image can be effectively improved by making the best of the license plate image characteristics, so as to obtain more accurate digital edge of the license plate, and provide license plate source images with higher quality to the extraction and analysis of the license plate information.

Description

Combine the MRF car plate Denoising Algorithm of priori based on gradient information and boxed area

Technical field

The invention belongs to Image Information Processing field is and in particular to combine priori based on gradient information and boxed area MRF car plate Denoising Algorithm.

Background technology

License plate image information is significant in safety monitoring, intelligent transportation and social security field, in overspeed of vehicle The aspect such as supervision, automobile burglar, highway and parking fee collective system, traffic accident monitoring plays very important effect.But low photograph The license plate image collecting under the conditions of degree often has the characteristics that signal to noise ratio is low, noise is big, noise contribution is complicated, to license board information Extract and analysis causes very big impact, cause Car license recognition to decline with following the tracks of accuracy rate.Therefore to noisy under low-light (level) environment Sound license plate image carries out effective denoising, improves license plate image quality, has important practical significance and actual application value.

Still less to car plate image denoising research under low-light (level) environment both at home and abroad at present.Under low-light (level) shooting condition, car plate Signal noise ratio (snr) of image is low, and noise intensity is big and complicated component, including Gaussian noise, poisson noise, impulsive noise, dark current noise Deng.Existing Denoising Algorithm often can only solve the relatively single situation of noise contribution it is difficult to adapt to noise in low-light (level) environment The situation of complicated component.

Markov random field (Markov Random Field, MRF) is one kind of probability graph model, applies at present In the field such as computer vision and image procossing, recover and reconstruction, estimation, texture analysis, edge inspection including image Survey, image segmentation, target identification and Attitude estimation etc..Its principle is to describe view data distribution, this condition using conditional probability Probability is unrelated with the position of pixel, and is depending on the relevant information in this field.Excavate car plate image prior letter abundant On the basis of breath, in conjunction with Markov random field model, build New Image recovery algorithms, be obtained in that more preferable license plate image Denoising effect.

Content of the invention

The present invention is to solve low-light (level) car plate Denoising Problems, provides one kind to combine priori based on gradient information and boxed area MRF car plate Denoising Algorithm, denoising and recovery can be carried out to the license plate image in noise video sequence.Implement including such as Lower step：

S1, utilize car plate digital gradient information, construct license plate image prior information, include gradient priori and boxed area elder generation Test；

S2, noise estimation is carried out to sequence of video images, determine noise profile probability density function；

S3, the two-value Markov random field model of structure license plate image, joint Markov random field model and car plate Image prior information, sets up optimization problem model and solves.

Further, described step S1 detailed process is：

(1) set I_vFor the car plate denoising image of vector form, the row, column number of image pixel is respectively m, n；Define I'_c, I'_r It is respectively column direction and line direction pixel value changes partial derivative matrix, for pixel (i, j), m, n, i, j are integer, its ladder Degree is expressed as lower form：

Wherein, G_(i,j)=[e_i+1+(j-1)m-e_i+(j-1)m,e_i+jm-e_i+(j-1)m]^T, e_kRepresent that kth position value is 1, other positions Put the mn dimensional vector that value is 0, k is integer, that is,：

By G_(i,j)It is combined into gradient matrix G：

License plate image gradient priori P_gFor：

Wherein, | | | |₂Represent and take 2 norm computings.

(2) calculate license plate image boxed area priori P_b：

N=mn is sum of all pixels in image, LP_bAnd LP_tIt is respectively car plate background area and character area pixel set, l_u For pixel P_uLabel value, l_vFor pixel P_vLabel value, belong to prospect or background, l for identifying pixel_uValue is：

In the same manner,

Further, described step S2 detailed process is：

If f^tFor the static background part of the corresponding video frame image of t, extract t- Δ t, during t- Δ t+1 ..., t+ Δ t Frame of video background parts { the f carving^t-Δt,…,f^t,…,f^t+Δt, Δ t represents time interval, is obtained approximately by frame accumulation method Noise-free picture

Wherein, η is variable, and span is t- Δ t, t- Δ t+1 ..., t+ Δ t；

Q frame background parts noise sample can be obtained by following formula：

Obtaining noise sample collection θ_ns={ n_q(q=t- Δ t, t- Δ t+1 ..., on the basis of t+ Δ t), utilizes Parzen windowhood method (the method is prior art, the method in particular reference [1]) obtains noise profile probability density Function：

Wherein, | θ_ns| represent noise sample collection θ_nsMiddle element number, kernel function W (x) form is：

Wherein, σ is the standard deviation of kernel function W (x), and x is function variable, and π is pi.

Further, described step S3 detailed process is：

If the noise-free picture recovering is I₀, the noise of (x, y) location of pixels is stochastic variable ε (x, y), (x, y) pixel Value i₁,i₂It is respectively background color value or prospect text color value, image plus noise process is expressed as：

I (x, y)=I₀(x,y)+ε(x,y),I₀(x,y)∈{i₁,i₂}

Image pixel value probability of occurrence is：

Wherein, the probability that p () occurs for event, δ (x, y) represents the neighborhood of pixel (x, y).Gone out with random field pixel value The weighted array of existing probability, car plate gradient priori and block priori builds the energy function of markov random file：

E_MRF=P_I+λ₁P_g+λ₂P_b

Wherein, P_IFor random field pixel value probability of occurrence, P_gFor license plate image gradient priori, P_bBlock first for license plate image Test, λ₁, λ₂For coefficient；

Image denoising problem can be attributed to solution following problems：

Wherein, l_u,l_v(u, v=1,2 ..., N) is the label value of denoising image pixel, I_vCar plate denoising for vector form Image, is asked to above-mentioned optimization using Graph Cuts algorithm (this algorithm is prior art, the method in particular reference [2]) Topic is solved, and finally obtains denoising image I^*.

The technique effect being obtained using the present invention：Compared with traditional car plate denoising method, this method can make full use of License plate image feature, effectively improves license plate image signal to noise ratio, obtains more accurate car plate digital edge, is that license board information is extracted There is provided higher-quality car plate source images with analysis.

Brief description

Fig. 1 is the flow chart of the method for the invention；

Fig. 2 is noiseless car plate source images；

Fig. 3 is low-light (level) noise license plate image；

Fig. 4 is that the present invention processes the license plate image obtaining；

Fig. 5 is the license plate image that BM3D algorithm obtains.

Specific embodiment

The invention will be further described with reference to the accompanying drawings and examples.

The invention provides a kind of MRF car plate Denoising Algorithm combining priori based on gradient information and boxed area, can Effective denoising and recovery are carried out to the license plate image in noise video sequence.As shown in figure 1, being the flow chart of the present invention, including Following steps：

Step one：Using car plate digital gradient information richness and block numeric area characteristic, construction is based on gradient and block The license plate image combination priori in shape region；

1. car plate gradient information information

If I_vFor the car plate denoising image of vector form, image pixel ranks number is respectively m, n.Define I'_c, I'_rIt is respectively Column direction and line direction pixel value changes partial derivative matrix.For pixel (i, j), its gradient approximately can be write as following shape Formula：

Wherein, G_(i,j)=[e_i+1+(j-1)m-e_i+(j-1)m,e_i+jm-e_i+(j-1)m]^T, e_kRepresent that kth position value is 1, other positions Put the mn dimensional vector that value is 0, that is,：

By G_(i,j)It is combined into gradient matrix G：

License plate image gradient priori P_gMay be calculated：

2. boxed area priori

Occur because license plate image numeral assumes boxed area, introduce boxed area priori P_b：

N=mn is sum of all pixels in image, LP_bAnd LP_tIt is respectively car plate background area and character area pixel set, l_u For pixel P_uLabel value, belong to prospect or background for identifying pixel.l_uValue is：

Step 2：By video noise estimation, determine license plate image noise profile probability density function.

If f^tStatic background part for the corresponding frame of video of t.Extract t- Δ t, t- Δ t+1 ..., t+ Δ t Frame of video background parts { f^t-Δt,…,f^t,…,f^t+Δt, approximate noise-free picture is obtained by frame accumulation method

Wherein q frame background parts noise sample can be obtained by following formula：

Obtaining noise sample collection θ_ns={ n_q(q=t- Δ t, t- Δ t+1 ..., on the basis of t+ Δ t), utilizes It is close that Parzen windowhood method (the method is prior art, specifically refers to the method in document [1]) obtains noise profile probability Degree function：

Wherein, | θ_ns| for noise sample collection θ_nsMiddle element number, kernel function W (x) form is：

Wherein, σ is the standard deviation of kernel function W (x).

Step 3：Joint Markov random field model and license plate image prior information, set up license plate image Denoising Problems Model simultaneously solves.

Noise license plate image is considered as two-dimentional markov random file, because license plate image has two-value, each pixel Value is background color value or prospect text color value.If needing the noise-free picture recovering to be I₀, the making an uproar of (x, y) location of pixels Sound is stochastic variable ε (x, y), value i of (x, y) pixel₁,i₂It is respectively background color value or prospect text color value.Image adds Noise process can be expressed as：

I (x, y)=I₀(x,y)+ε(x,y),I₀(x,y)∈{i₁,i₂}

Pixel probability is：

With the weighted array of random field pixel value probability of occurrence, car plate gradient priori and block priori build Markov with The energy function on airport：

E_MRF=P_I+λ₁P_g+λ₂P_b

Wherein, P_IFor image random field pixel value probability of occurrence, P_gFor license plate image gradient priori, P_bFor license plate image block Shape priori, λ₁, λ₂For coefficient.Image denoising problem can be attributed to solution following problems：

Wherein, l_u,l_v(u, v=1,2 ..., N) is the label value of denoising image pixel, I_vCar plate denoising for vector form Image.

For the car plate source images shown in Fig. 2, its corresponding single frames low-light (level) noise image is as shown in figure 3, adopt above-mentioned Method sets up Image restoration, and using Graph Cuts algorithm, (this algorithm is prior art, in particular reference [2] Method) optimization problem in step 3 is solved, obtain image result as shown in Figure 4.

Process the result obtaining using BM3D algorithm (this algorithm is prior art, the method in particular reference [3]) As shown in Figure 5；Fig. 4 is compared with Fig. 5, draw application this method can recover to become apparent from, the more preferable car of degree of identification Board digital picture.The bibliography being related in literary composition is as follows：

[1]T.Hastie,R.Tibshirani,and J.Friedman.The elements of statistical learning:Data mining,inference,and prediction.Springer,2001.

[2]V.Kolmogorov and R.Zabih.What energy functions can be minimized via graph cuts？IEEE TPAMI,2004.

[3]Dabov K,Foi A,Katkovnik V,et al.Image denoising by sparse 3D transform-domain collaborative filtering.IEEE Transactions on Image Processing,2007.

The above is only the preferred embodiment of the present invention, protection scope of the present invention is not limited merely to above-described embodiment, All technical schemes belonging under thinking of the present invention belong to protection scope of the present invention.It should be pointed out that for the art For those of ordinary skill, some improvements and modifications without departing from the principles of the present invention, should be regarded as the protection of the present invention Scope.

Claims (1)

1. a kind of MRF car plate Denoising Algorithm combining priori based on gradient information and boxed area, MRF represents that markov is random Field is it is characterised in that comprise the steps：

S1, utilize car plate digital gradient information, construct license plate image prior information, including gradient priori and boxed area priori； Detailed process is：

(1) set I_vFor the car plate denoising image of vector form, image pixel ranks number is respectively m, n；Define I'_c, I'_rIt is respectively Column direction and line direction pixel value changes partial derivative matrix, for pixel (i, j), its gradient is expressed as lower form：

$G_{(i,j)}{I}_v=\left(\begin{array}{c}{I}_{c(i,j)}^{\′}\\ I_{r(i,j)}^{\′}\end{array}\right)$

Wherein, G_(i,j)=[e_i+1+(j-1)m-e_i+(j-1)m,e_i+jm-e_i+(j-1)m]^T, e_kRepresent that kth position value is that 1, other positions take It is worth the mn dimensional vector for 0, that is,：

By G_(i,j)It is combined into gradient matrix G：

$G=\left(\begin{array}{c}{G}_{(1,1)}\\ .\\ .\\ .\\ G_{(m,n)}\end{array}\right)$

License plate image gradient priori P_gFor：

$P_g=\underset{i=1}{\overset{m}{\Σ}}\underset{j=1}{\overset{n}{\Σ}}\left|\right|G_{(i,j)}{I}_v|{|}_2$

Wherein, | | | |₂Represent and take 2 norm computings；

(2) calculate license plate image boxed area priori P_b：

$P_b=\exp (\underset{u=1}{\overset{N}{\Σ}}\underset{v=1}{\overset{N}{\Σ}}{l}_u{l}_v+(1-l_u\left)\right(1-l_v\left)\right),(u,v=1,2,\mathrm{...},N)$

N=mn is sum of all pixels in image, LP_bAnd LP_tIt is respectively car plate background area and character area pixel set, l_uIt is picture Plain P_uLabel value, belong to prospect or background, l for identifying pixel_uValue is：

$l_u=\left\{\begin{array}{cc}0,& P_u\∈{\mathrm{LP}}_b\\ 1,& P_u\∈{\mathrm{LP}}_t\end{array}\right.,(u=1,2,\mathrm{...},N)$

S2, noise estimation is carried out to sequence of video images, determine noise profile probability density function；Detailed process is：

If f^tFor the static background part of the corresponding video frame image of t, extract t- Δ t, t- Δ t+1 ..., the regarding of t+ Δ t Frequency frame background parts { f^t-Δt,…,f^t,…,f^t+Δt, Δ t represents time interval, obtains noise-free picture by frame accumulation method

$\stackrel{\‾}{f}=\frac{1}{2\mathrm{\Δ}t+1}\underset{\mathrm{\η}=t-\mathrm{\Δ}t}{\overset{t+\mathrm{\Δ}t}{\Σ}}{f}^{\mathrm{\η}}$

Wherein, η is variable, and span is t- Δ t, t- Δ t+1 ..., t+ Δ t；

Q frame background parts noise sample can be obtained by following formula：

$n_q=f^q-\stackrel{\‾}{f},(q=t-\mathrm{\Δ}t,t-\mathrm{\Δ}t+1,\mathrm{...},t+\mathrm{\Δ}t)$

Obtaining noise sample collection θ_ns={ n_q(q=t- Δ t, t- Δ t+1 ..., on the basis of t+ Δ t), calculates and obtains noise Distribution probability density function：

$f_{pn}\left(z\right)=\frac{1}{\left|{\mathrm{\θ}}_{ns}\right|}\underset{j=1}{\overset{\left|{\mathrm{\θ}}_{ns}\right|}{\Σ}}W(z-n_j)$

Wherein, | θ_ns| represent noise sample collection θ_nsMiddle element number, kernel function W (x) form is：

$W\left(x\right)=\frac{1}{\mathrm{\σ}\sqrt{2\mathrm{\π}}}{e}^{-x^2/2{\mathrm{\σ}}^2}$

Wherein, σ is the standard deviation of kernel function W (x)；

S3, the two-value Markov random field model of structure license plate image, joint Markov random field model and license plate image Prior information, sets up optimization problem model and solves, and detailed process is：

If the noise-free picture recovering is I₀, the noise of (x, y) location of pixels is stochastic variable ε (x, y), the value of (x, y) pixel i₁,i₂It is respectively background color value or prospect text color value, image plus noise process is expressed as：

I (x, y)=I₀(x,y)+ε(x,y),I₀(x,y)∈{i₁,i₂}

Image pixel value probability of occurrence is：

$p\left(I\right|I_0\left(x,y\right))=\underset{(x^{\′},y^{\′})\∈\mathrm{\δ}(x,y)}{\Π}\left(f_{pn}\right(I\left(x^{\′},y^{\′}\right)-I_0\left(x^{\′},y^{\′}\right)\left)\right)$

Wherein, δ (x, y) is the neighborhood of pixel (x, y), with random field pixel value probability of occurrence, car plate gradient priori and bulk first The weighted array tested builds the energy function of markov random file：

E_MRF=P_I+λ₁P_g+λ₂P_b

Wherein, P_IFor image random field pixel value probability of occurrence, P_gFor license plate image gradient priori, P_bBlock first for license plate image Test, λ₁, λ₂For coefficient；

Image denoising problem can be attributed to solution following problems：

$I^{*}=\underset{I}{\arg \max }{E}_{MRF}$

$\left\{\begin{array}{c}{P}_I=\underset{u=1}{\overset{N}{\mathrm{\Π}}}\frac{p(l_u=0)}{p(l_u=1)}\\ P_g=\underset{i=1}{\overset{m}{\Σ}}\underset{j=1}{\overset{n}{\Σ}}\left|\right|G_{(i,j)}{I}_v|{|}_2\\ P_b=\mathrm{\λ}\exp (\underset{u=1}{\overset{N}{\Σ}}\underset{v=1}{\overset{N}{\Σ}}{l}_u{l}_v+(1-l_u)(1-l_v))\end{array}\right.$

Wherein, l_u,l_v(u, v=1,2 ..., N) is the label value of denoising image pixel, I_vCar plate denoising figure for vector form Picture, solves to above formula optimization problem, finally obtains denoising image I^*.