CN106651804A - Vehicle-mounted image de-noising method and system based on adaptive diffusion filtering - Google Patents

Abstract

The invention provides a vehicle-mounted image de-noising method and system based on adaptive diffusion filtering, and the method comprises the steps: 1, inputting a noise image, and carrying out the extension of the boundary of the noise image; 2, selecting a filter group, and constructing a diffusion flux function; 3, carrying out the loop iteration based on explicit forwarding difference, and starting the image adaptive diffusion; 4, cutting the boundary of the de-noised image, and obtaining the de-noised image with the same size as an original inputted noise image. The beneficial effects of the invention are that the method employs the improved diffusion flux function, and can cause the adaptive forward/backward diffusion behaviors, so the method can give consideration to the filtering and protection and enhancement of the edge and texture of a flat region of the image.

Description

Vehicle-mounted image de-noising method and system based on self adaptation diffusing filter

Technical field

The present invention relates to technical field of image processing, more particularly to the vehicle-mounted image denoising side based on self adaptation diffusing filter Method and system.

Background technology

With the enhancing of people's awareness of safety, and to driving the raising that comfort level is required, senior drive assist system (Advanced Driver Assistance System, ADAS) increases substantially occurring in that in recent years.In a variety of systems, DAS (Driver Assistant System) city based on image accounts for rate highest.It is with low cost that its main cause is, and can be with drive recorder It is used in combination, and the result of detecting can be presented to driver in the way of vision imaging, thus it is very popular.However, vehicle-mounted The gatherer process of image is often disturbed (as light illumination condition is poor) by noise, so that the quality of vehicle-mounted image is greatly Decline.Picture noise affects follow-up higher level image analysis tasks, such as row except affecting visual experience, more seriously People, lane detection, Target Segmentation etc..In order to ensure the visibility and reliability of the DAS (Driver Assistant System) of view-based access control model, one is found Feasible vehicle-mounted image de-noising method is planted, enhancing process is carried out to DAS (Driver Assistant System) image and is had great importance.

Existing image de-noising method includes filter in spatial domain (such as medium filtering, bilateral filtering, non-local mean filtering Deng), frequency filtering (such as Fourier transformation, wavelet transformation), the denoising method (such as K-SVD) based on rarefaction representation and Denoising method (such as P-M models) based on Anisotropic diffusion.Wherein filter in spatial domain method is typically relatively simple, thus effect Fruit is slightly poor.[S.Gu, L.Zhang, W.Zuo, and X.Feng, " Weighted Nuclear Norm Minimization in the recent period With Application to Image Denoising, " In CVPR 2014] propose a kind of improved non local filtering Method-WNNM.It utilizes the non local similitude of image and matrix rank minimization, obtain one stream denoising effect, but the time Complexity is of a relatively high.Frequency domain filtering while high-frequency noise is filtered, to the edge with high frequency characteristics and texture structure Destroy, reduced the quality of denoising image.Dictionary learning process operand based on sparse representation method is larger.Compare it Under, Anisotropic diffusion model has (1) simple structure, and time complexity is low；(2) can preferably take into account noise removal and Advantage of both the holding of the details such as edge, texture, thus the always focus of research.

Common image diffusion model typically stems from the P-M models that Perona and Malik are proposed in nineteen ninety.With anti- Answer the discrete form such as formula (1) of the P-M models of item shown

Wherein, u⁰For initial noisc image, * is that two-dimensional convolution is operated, k_i* u represents image u and linear filter k_iVolume Product,Represent wave filter k_i180 degree is rotated around central point, φ is the flux function (flux for controlling dispersal behavior Function), it is usually taken to be

K is threshold parameter, and λ is the intensity for reacting item.Wave filter k in formula (1)_xAnd k_yRespectively

Image and wave filter k_xConvolution be gradient of the image in x directions, image and wave filter k_yConvolution be image in y The gradient in direction.

After proposing from P-M models, anisotropy parameter technology has obtained extensive concern and research.(Chinese patent CN 105427262) adaptive design and improvement have been carried out to Grads threshold so as to according to the maximum gradation value and iterations of image Automatically control Grads threshold.(Chinese patent CN 101877122) discloses a kind of anisotropic filtering based on trace operator model Method, this is a kind of Anisotropic Diffusion Model based on local geometry.(Chinese patent CN 104166965) is disclosed A kind of method that entropy based on gradient magnitude selects diffusion flux function.

As the research to Anisotropic diffusion model deepens continuously, test result indicate that, also there is following lacking in the technology Point：

Firstth, due to wave filter k that diffusion process is used_xAnd k_yOnly comprising the correlation between nearest neighbor pixels, cause There is significantly " ladder " effect in image after denoising, that is, the region of a large amount of piecemeal constant brightness values occur；

Secondth, because the flux function of the control dispersal behavior for being adopted is only capable of slowing down in regions such as image border, textures The speed of image diffusion, can not make diffusion stop or even carry out counter diffusion (i.e. image border enhancing), so as to cause to protect edge Property is not fine.

A kind of Image denoising algorithm how is designed, simple structure is facilitated implementation, and can simultaneously reach picture noise removal Protection and enhanced purpose with the detailed information such as edge, texture, improves denoising performance, is urgent need to resolve in vehicle-mounted image procossing Problem.

The content of the invention

The invention provides a kind of vehicle-mounted image de-noising method based on self adaptation diffusing filter, comprises the steps：

Step one：Input noise image, to the border of noise image continuation is carried out；

Step 2：Selected wave filter group, constructs diffusion flux function；

Step 3：Selected wave filter group, constructs diffusion flux function；

Step 4：Cutting is carried out to denoising image boundary, the equivalently-sized denoising figure of noise image is obtained and be originally inputted Picture.

As a further improvement on the present invention, in the step 2, using linear filter diffusion flux function is constructed.

As a further improvement on the present invention, in the step 2：

Selected linear filter group F={ f₁,f₂,f₃…f_N, include N number of wave filter f_i(i=1～N)；Build diffusion logical Flow function

Wherein, K_f, K_b, W, α are the parameter of control function shape, build diffusion equation

Wherein f_iFor i-th element of wave filter group F,Represent wave filter f_i180 degree is rotated around central point, * is two dimension volume Product operation, θ_iFor the weighted value of i-th filter results, λ is the weight for reacting item, θ_iPositive number is with λ.

As a further improvement on the present invention, comprise the steps in the step 3：

(1) for the image u of t^t, it is calculated with linear filter f_iConvolution, i.e. u^t*f_i；

(2) by diffusion flux function phi₂Z () is applied to u according to mode pixel-by-pixel^t*f_iResult on, obtain φ₂(u^t* f_i)；

(3) φ is calculated₂(u^t*f_i) and wave filterThe result of convolution simultaneously considers weighted value θ_iObtain

(4) above-mentioned steps 1) -3) complete to a linear filter f_iCalculating, repeating n times can complete to wave filter The calculating of all wave filters in group F, summation can be obtained

(5) calculate reaction and expand item λ (u^t-u⁰)；

(6) rule is updated using iteration

Calculate u^t+1, wherein Δ t is the time step of diffusion process；

(7) judge whether to complete iteration, if, then obtain spreading result, otherwise return execution step (1).

As a further improvement on the present invention, the linear filter is that the two-dimension discrete cosine transform that size is 7 × 7 is filtered Ripple device group.

Present invention also offers a kind of vehicle-mounted image denoising system based on self adaptation diffusing filter, including：

Input module：For input noise image, continuation is carried out to the border of noise image；

Constructing module：For selecting wave filter group, diffusion flux function is constructed；

Processing module：For being circulated iteration based on explicit forward difference, image adaptive diffusion starts；

Output module：Cutting is carried out to denoising image boundary, the equivalently-sized denoising of noise image is obtained and be originally inputted Image.

As a further improvement on the present invention, in the constructing module, using linear filter diffusion flux letter is constructed Number.

As a further improvement on the present invention, in the constructing module：

Selected linear filter group F={ f₁,f₂,f₃…f_N, include N number of wave filter f_i(i=1～N)；Build diffusion logical Flow function

Wherein, K_f, K_b, W, α are the parameter of control function shape, build diffusion equation

Wherein f_iFor i-th element of wave filter group F,Represent wave filter f_i180 degree is rotated around central point, * is two dimension volume Product operation, θ_iFor the weighted value of i-th filter results, λ is the weight for reacting item, θ_iPositive number is with λ.

As a further improvement on the present invention, include in the processing module：

First processes submodule：For the image u of t^t, it is calculated with linear filter f_iConvolution, i.e. u^t*f_i；

Second processing submodule：By diffusion flux function phi₂Z () is applied to u according to mode pixel-by-pixel^t*f_iResult On, obtain φ₂(u^t*f_i)；

3rd processes submodule：Calculate φ₂(u^t*f_i) and wave filterThe result of convolution simultaneously considers weighted value θ_iObtain

Fourth process submodule：Completed to a linear filter f to the 3rd process submodule by the first process submodule_i Calculating, repeat the calculating that n times can complete to all wave filters in wave filter group F, summation can be obtained

5th processes submodule：Calculate reaction and expand item λ (u^t-u⁰)；

6th processes submodule：Rule is updated using iteration

Calculate u^t+1, wherein Δ t is the time step of diffusion process；

Judge module：Judge whether to complete iteration, if, then obtain spreading result, otherwise return and perform the first process Submodule.

As a further improvement on the present invention, the linear filter is that the two-dimension discrete cosine transform that size is 7 × 7 is filtered Ripple device group.

The invention has the beneficial effects as follows：Present invention uses improved diffusion flux function, can result in it is adaptive before To/backward dispersal behavior, therefore, it is possible to take into account the protection filtered with the structural information such as edge, texture of image flat site noise And enhancing.

Description of the drawings

Fig. 1 is the flow chart of the present invention.

Fig. 2 is the curve comparison figure of diffusion flux function of the present invention and classical flux function, φ₁Z () is normal Diffusion flux function, φ₂Z () is flux function of the present invention.

Fig. 3 is the DCT wave filter groups that one embodiment of the invention is adopted, and filter size is 7 × 7.

Fig. 4 is the weighted value figure of the DCT wave filter groups correlation that one embodiment of the invention is adopted.

Fig. 5 is the experiment test image of the present invention.

Fig. 6 is denoising effect figure of the different denoising methods to image man.

Specific embodiment

As shown in figure 1, the invention discloses a kind of vehicle-mounted image de-noising method based on self adaptation diffusing filter, including such as Lower step：

Step one：Input noise image u⁰, continuation is carried out to the border of noise image, to prevent follow-up image from spreading behaviour Make to produce artifact at image boundary, may be used to order in Matlab and complete, continuation size and subsequent filter size phase Together；

Step 2：Selected wave filter group, constructs diffusion flux function；

Step 3：Iteration is circulated based on explicit forward difference, image adaptive diffusion starts；

Step 4：Cutting is carried out to denoising image boundary, the equivalently-sized denoising figure of noise image is obtained and be originally inputted Picture.

In step 2, linear filter group F={ f is selected₁,f₂,f₃…f_N, include N number of wave filter f_i(i=1～ N)；Build diffusion flux function

Wherein, K_f, K_b, W, α are the parameter of control function shape, build diffusion equation

Wherein f_iFor i-th element of wave filter group F,Represent wave filter f_i180 degree is rotated around central point, * is two dimension volume Product operation, θ_iFor the weighted value of i-th filter results, λ is the weight for reacting item, θ_iPositive number is with λ.

In step 3, comprise the steps：

(1) for the image u of t^t, it is calculated with linear filter f_iConvolution, i.e. u^t*f_i, convolutional calculation can adopt Symmetrical boundary condition, in Matlab to order can realize imfilter (u, f, ' symmetric', ' conv ')；；

(2) by diffusion flux function phi₂Z () is applied to u according to mode pixel-by-pixel^t*f_iResult on, obtain φ₂(u^t* f_i)；

(3) φ is calculated₂(u^t*f_i) and wave filterThe result of convolution simultaneously considers weighted value θ_iObtain Equally can be completed with Matlab order imfilter；

(4) above-mentioned steps 1) -3) complete to a linear filter f_iCalculating, repeating n times can complete to wave filter The calculating of all wave filters in group F, summation can be obtained

(5) calculate reaction and expand item λ (u^t-u⁰)；

(6) rule is updated using iteration

Calculate u^t+1, wherein Δ t is the time step of diffusion process, if Δ t=0.2；

(7) judge whether to complete iteration, if, then obtain spreading result, otherwise return execution step (1).

Above-mentioned steps (1)-(6) complete an iteration, and repeatedly iteration is obtained picture rich in detail, for example, repeatedly change for 50 times In generation, is obtained picture rich in detail.

As a preferred embodiment of the present invention, the two-dimension discrete cosine transform wave filter group that size is 7 × 7 has been used (Discrete Cosine Transform, DCT), its 48 wave filter deducted outside constant value wave filter are shown in accompanying drawing 3, i-th DCT wave filters are designated as b_i, b_iMould a length of 1.

If wave filter f_i=β_i·b_i

I-th wave filter f_iThe weighted value β of association_iAnd θ_iSee accompanying drawing 4, the weight for reacting item is set to λ=1.Diffusion flux letter Number φ₂Z the parameter setting of () is K_f=4, K_b=20, W=8, α=0.05, its curve map is shown in accompanying drawing 2.

Furthermore, it is to be understood that the selection of wave filter group can have many forms, such as PCA and Gabor filter group.Meanwhile, The selection of diffusion flux function parameter also has various ways, specifically can be set according to actual application environment.

The invention also discloses a kind of vehicle-mounted image denoising system based on self adaptation diffusing filter, including：

Input module：For input noise image, continuation is carried out to the border of noise image；

Constructing module：For selecting wave filter group, diffusion flux function is constructed；

Processing module：For being circulated iteration based on explicit forward difference, image adaptive diffusion starts；

Output module：Cutting is carried out to denoising image boundary, the equivalently-sized denoising of noise image is obtained and be originally inputted Image.

In the constructing module, using linear filter diffusion flux function is constructed.

In the constructing module：

Selected linear filter group F={ f₁,f₂,f₃…f_N, include N number of wave filter f_i(i=1～N)；Build diffusion logical Flow function

Wherein, K_f, K_b, W, α are the parameter of control function shape, build diffusion equation

Wherein f_iFor i-th element of wave filter group F,Represent wave filter f_i180 degree is rotated around central point, * is two dimension volume Product operation, θ_iFor the weighted value of i-th filter results, λ is the weight for reacting item, θ_iPositive number is with λ.

Include in the processing module：

First processes submodule：For the image u of t^t, it is calculated with linear filter f_iConvolution, i.e. u^t*f_i；

Second processing submodule：By diffusion flux function phi₂Z () is applied to u according to mode pixel-by-pixel^t*f_iResult On, obtain φ₂(u^t*f_i)；

3rd processes submodule：Calculate φ₂(u^t*f_i) and wave filterThe result of convolution simultaneously considers weighted value θ_iObtain

Fourth process submodule：Completed to a linear filter f to the 3rd process submodule by the first process submodule_i Calculating, repeat the calculating that n times can complete to all wave filters in wave filter group F, summation can be obtained

5th processes submodule：Calculate reaction and expand item λ (u^t-u⁰)；

6th processes submodule：Rule is updated using iteration

Calculate u^t+1, wherein Δ t is the time step of diffusion process；

Judge module：Judge whether to complete iteration, if, then obtain spreading result, otherwise return and perform the first process Submodule.

Patent of the present invention has been substantially carried out two aspects on the basis of traditional anisotropic image diffusion model Improve：

1. than conventional First-order Gradient wave filter, present invention uses the linear filter of higher order.The chi of wave filter Very little bigger, the neighborhood of pixels of covering is also bigger, thus can preferably extract Local Structure of Image information, and then can preferably keep Even strengthen the partial structurtes such as image texture, edge；

2. used a kind of improved flux function to control image dispersal behavior, its control with conventional flux function Curve map such as Fig. 2.This flux function can result in before adaptive image to diffusion (image smoothing) and backward diffusion (image Strengthen) process, the protection and enhancing that filter with the structural information such as edge, texture of image flat site noise can be taken into account.

Flux function φ used in the present invention₂(z) and conventional flux function φ₁Z the key difference between () is, It has the flex point of a zero crossing, A points as shown in Figure 2.Flux function is the function with regard to wave filter response.In φ₂ In (z), when the response of wave filter is smaller, (image flat site is corresponded to) at the left side of flex point, flux function for just, Corresponding image dispersal behavior is positive diffusion (image smoothing)；When the response of wave filter is than larger, more than (correspondence during flex point For image border, texture region), flux function is negative, and corresponding image dispersal behavior is reverse diffusion (image enhaucament).

Due to above-mentioned both sides reason, the present invention has the advantages that following：

1., due to containing the neighborhood information of higher order, the present invention can solve the problem that " ladder " effect；

2. present invention uses improved diffusion flux function, can result in adaptive forward direction/backward dispersal behavior, because This can take into account the protection and enhancing that filter with the structural information such as edge, texture of image flat site noise；

3. the present invention have the advantages that conventional diffusion model structure simply, be easily achieved, time complexity it is low, while defeated Going out signal to noise ratio and the aspect of visual effect two can also reach first-class Denoising Algorithm, the such as performance (see accompanying drawing 6) of WNNM algorithms.

The effect of the present invention can be further characterized by by following experiment.Inventive algorithm is carried out under Matlab environment Emulation, and be compared with the WNNM algorithms of classics P-M algorithms and denoising effect one stream.The objective evaluation index of denoising effect With Y-PSNR (Peak Signal-Noise Ratio, PSNR) and structural similarity (Structural SIMilarity, SSIM) weigh.

Experiment condition：On PC containing 4 Intel CPUs and 8GB RAM, under 2015 editions Matlab programmed environments.Experiment Shown in the input picture for being used such as Fig. 5 (a, b), Gaussian noise is firstly added.Add the image such as Fig. 5 after white Gaussian noise Shown in (c, d), standard deviation sigma=25 of noise, the Y-PSNR PSNR=20.18 of the image after plus noise, SSIM= 0.3303。

Experiment content：Under these experimental conditions, it is first-class from P-M models (classic map is as diffusion model) and at present WNNM denoising methods and the inventive method carry out Experimental comparison.The objective evaluation index of denoising effect with Y-PSNR PSNR and Structural similarity SSIM is measured.

The adopted image of experiment contains a large amount of texture regions, such as the feather accessories and the fold of clothes of cap.Three The denoising effect of the method for kind is shown in accompanying drawing 6.From the point of view of visual effect, the image after WNNM algorithms and the inventive method denoising exists These texture regions all achieve the visual effect for making us relatively satisfactory, and texture structure has obtained preferable holding, and denoising result is seen To get up also compare nature.And P-M models have been greatly lowered denoising image just like obvious " ladder " effect is expectedly generated Visual effect, while objective evaluation index is also than relatively low, it is as shown in the table.

Following table gives the SSIM values and PSNR values of three kinds of denoising methods and compares.

Evaluation index	Plus noise image	P-M models	WNNM algorithms	Inventive algorithm
					SSIM	0.3303	0.7357	0.8047	0.8114
PSNR	20.18	28.29	29.62	29.88

As can be seen from the above table, the inventive method is similar to WNNM in two objective evaluation indexs, than classical P-M Model will get well.Comprehensive visual effect and the aspect of objective evaluation standard two, denoising method proposed by the present invention is than traditional image Diffusion model has obvious superiority.

Above content is to combine specific preferred embodiment further description made for the present invention, it is impossible to assert The present invention be embodied as be confined to these explanations.For general technical staff of the technical field of the invention, On the premise of without departing from present inventive concept, some simple deduction or replace can also be made, should all be considered as belonging to the present invention's Protection domain.

Claims (10)

1. a kind of vehicle-mounted image de-noising method based on self adaptation diffusing filter, it is characterised in that comprise the steps：

Step one：Input noise image, to the border of noise image continuation is carried out；

Step 2：Selected wave filter group, constructs diffusion flux function；

Step 3：Iteration is circulated based on explicit forward difference, image adaptive diffusion starts；

Step 4：Cutting is carried out to denoising image boundary, the equivalently-sized denoising image of noise image is obtained and be originally inputted.

2. vehicle-mounted image de-noising method according to claim 1, it is characterised in that in the step 2, using linear Wave filter constructs diffusion flux function.

3. vehicle-mounted image de-noising method according to claim 2, it is characterised in that in the step 2：

Selected linear filter group F={ f₁,f₂,f₃…f_N, include N number of wave filter f_i(i=1～N)；Build diffusion flux letter Number

${\mathrm{\φ}}_2\left(z\right)=\frac{z}{1+{\left(\frac{z}{K_f}\right)}^4}-\frac{\mathrm{\α}z}{1+{\left(\frac{z-K_b}{W}\right)}^4}$

Wherein, K_f, K_b, W, α are the parameter of control function shape, build diffusion equation

$\frac{\∂u}{\∂t}=-\underset{i=1}{\overset{N}{\Σ}}{\mathrm{\θ}}_i\·{\stackrel{\‾}{f}}_i*{\mathrm{\φ}}_2(f_i*u)-\mathrm{\λ}(u-u^0)---\left(2\right)$

Wherein f_iFor i-th element of wave filter group F,Represent wave filter f_i180 degree is rotated around central point, * is two-dimensional convolution behaviour Make, θ_iFor the weighted value of i-th filter results, λ is the weight for reacting item, θ_iPositive number is with λ.

4. vehicle-mounted image de-noising method according to claim 2, it is characterised in that include following step in the step 3 Suddenly：

(1) for the image u of t^t, it is calculated with linear filter f_iConvolution, i.e. u^t*f_i；

(2) by diffusion flux function phi₂Z () is applied to u according to mode pixel-by-pixel^t*f_iResult on, obtain φ₂(u^t*f_i)；

(3) φ is calculated₂(u^t*f_i) and wave filterThe result of convolution simultaneously considers weighted value θ_iObtain

(4) above-mentioned steps 1) -3) complete to a linear filter f_iCalculating, repeating n times can complete in wave filter group F The calculating of all wave filters, summation can be obtained

(5) calculate reaction and expand item λ (u^t-u⁰)；

(6) rule is updated using iteration

$u^{t+1}=u^t-\mathrm{\Δ}t\·(\underset{i=1}{\overset{N}{\Σ}}{\mathrm{\θ}}_i\·{\stackrel{\‾}{f}}_i*{\mathrm{\φ}}_2(f_i*u^t)-\mathrm{\λ}(u^t-u^0\left)\right)---\left(3\right)$

Calculate u^t+1, wherein Δ t is the time step of diffusion process；

(7) judge whether to complete iteration, if, then obtain spreading result, otherwise return execution step (1).

5. the vehicle-mounted image de-noising method according to any one of claim 2 to 4, it is characterised in that the linear filter Be size be 7 × 7 two-dimension discrete cosine transform wave filter group.

6. a kind of vehicle-mounted image denoising system based on self adaptation diffusing filter, it is characterised in that include：

Input module：For input noise image, continuation is carried out to the border of noise image；

Constructing module：For selecting wave filter group, diffusion flux function is constructed；

Processing module：For being circulated iteration based on explicit forward difference, image adaptive diffusion starts；

Output module：Cutting is carried out to denoising image boundary, the equivalently-sized denoising image of noise image is obtained and be originally inputted.

7. vehicle-mounted image denoising system according to claim 6, it is characterised in that in the constructing module, using line Property wave filter construction diffusion flux function.

8. vehicle-mounted image denoising system according to claim 7, it is characterised in that in the constructing module：

Selected linear filter group F={ f₁,f₂,f₃…f_N, include N number of wave filter f_i(i=1～N)；Build diffusion flux letter Number

${\mathrm{\φ}}_2\left(z\right)=\frac{z}{1+{\left(\frac{z}{K_f}\right)}^4}-\frac{\mathrm{\α}z}{1+{\left(\frac{z-K_b}{W}\right)}^4}$

Wherein, K_f, K_b, W, α are the parameter of control function shape, build diffusion equation

$\frac{\∂u}{\∂t}=-\underset{i=1}{\overset{N}{\Σ}}{\mathrm{\θ}}_i\·{\stackrel{\‾}{f}}_i*{\mathrm{\φ}}_2(f_i*u)-\mathrm{\λ}(u-u^0)---\left(2\right)$

Wherein f_iFor i-th element of wave filter group F,Represent wave filter f_i180 degree is rotated around central point, * is two-dimensional convolution behaviour Make, θ_iFor the weighted value of i-th filter results, λ is the weight for reacting item, θ_iPositive number is with λ.

9. vehicle-mounted image denoising system according to claim 7, it is characterised in that include in the processing module：

First processes submodule：For the image u of t^t, it is calculated with linear filter f_iConvolution, i.e. u^t*f_i；

Second processing submodule：By diffusion flux function phi₂Z () is applied to u according to mode pixel-by-pixel^t*f_iResult on, obtain To φ₂(u^t*f_i)；

3rd processes submodule：Calculate φ₂(u^t*f_i) and wave filterThe result of convolution simultaneously considers weighted value θ_iObtain

Fourth process submodule：Completed to a linear filter f to the 3rd process submodule by the first process submodule_i's Calculate, the calculating to all wave filters in wave filter group F is completed by repeating n times, summation can be obtained

5th processes submodule：Calculate reaction and expand item λ (u^t-u⁰)；

6th processes submodule：Rule is updated using iteration

$u^{t+1}=u^t-\mathrm{\Δ}t\·(\underset{i=1}{\overset{N}{\Σ}}{\mathrm{\θ}}_i\·{\stackrel{\‾}{f}}_i*{\mathrm{\φ}}_2(f_i*u^t)-\mathrm{\λ}(u^t-u^0\left)\right)---\left(3\right)$

Calculate u^t+1, wherein Δ t is the time step of diffusion process；

Judge module：Judge whether to complete iteration, if, then obtain spreading result, otherwise return and perform the first process submodule Block.

10. the vehicle-mounted image denoising system according to any one of claim 7 to 9, it is characterised in that the linear filter Be size be 7 × 7 two-dimension discrete cosine transform wave filter group.