CN103927723A - Image filtering method based on neuro-fuzzy system and edge detection - Google Patents

Abstract

The invention relates to an image filtering method based on a neuro-fuzzy system and edge detection. For an image polluted by impulse noise, the method combines a median filter, an edge detector and a self-adaptation neuro-fuzzy inference system, before the neuro-fuzzy system is used for carrying out filtering on the noise image, a training image is constructed artificially at first, a blended learning algorithm is used for training the system and determining parameters in the system, and then the optimized system is used for carrying out noise filtering on the image polluted by the impulse noise. The image filtering method based on the neuro-fuzzy system and edge detection can effectively filter away the impulse noise in the image and well reserve edges and details in the original image, and the filtering performance of the image filtering method is better than that of other traditional image filtering methods.

Description

Based on the image filtering method of fuzzy neuron system and rim detection

Technical field

The present invention relates to the image filtering method based on fuzzy neuron system and rim detection, belong to technical field of image processing, be specifically related to a kind of spiced salt impulsive noise image filtering method.

Background technology

Digital picture is obtained, may be caused because of various reasons in transmitting procedure sneaking into noise signal in normal view data at it.These noise signals have reduced the quality of digital picture, have also affected the exploitation of subsequent applications.For addressing this problem, association area researchist has proposed the method for many image denoisings.

In image,, also there are abundant details or edge in general existing comparatively level and smooth region, and these details or edge comprise important visually-perceptible information conventionally.Therefore, the object of image filtering, except removing noise, also will retain the information such as details or edge as much as possible.In the image filtering technology for being polluted by impulsive noise, the effect of nonlinear filtering is better than linear filtering, because linear filtering can cause details and the edge fog of image, affects the visual effect of image.Standard median filter (standard median filter, SMF) method is a kind of non-linear filtering method of classics, and it replaces the gray-scale value of this window center point pixel by the intermediate value of the gray-scale value of all pixels in predefined filter window.For further improving the filtering performance of SMF, O.Y1i-Harja etc. have proposed weighted median filter (weighted median filter, WMF), S.J.Ko etc. have proposed center weighted median filter (center weighted median filter, CWMF), this two classes wave filter gives the specific more weights of pixel in filtering operation window.

These three kinds of wave filters all adopt identical filtering processing for pixels all in noise image above, and this just inevitably destroys not by the pixel of noise pollution in filtering.In actual applications, we wish that filtering algorithm only carries out filtering to noise pixel, remains unchanged to non-noise pixel.Therefore, before filtering, can adopt an impulse noise detection device, the noise pixel in image and non-noise pixel are made a distinction.If it is noise pixel that a pixel is detected, it will be substituted by the output of median filter; Otherwise this pixel remains unchanged.According to adopting different impulse noise correction method, Z.Shuqun etc. have proposed edge-detecting median filter (EDMF) method, C.Tao etc. have proposed multi-state median filter (MSMF) method, E.Abreu etc. have proposed signal-dependent rand-ordered mean filter (SDROMF) method, and Z.Wang etc. have proposed progressive switching median filter (PSMF) method.

After image is by noise pollution, make wave filter be difficult to accurately distinguish the difference between details or edge and noise, and in filtering, can inevitably there is uncertainty and the imperfection of information, and fuzzy theory just in time can adapt to these uncertainty, therefore in image filtering process, apply fuzzy theory and can obtain good noise remove effect.F.Russo etc. have utilized fuzzy system theory just, have proposed fuzzy filter (FF) method, are applied in image filtering.Artificial neural network algorithm shows very large superiority compared with traditional algorithm, artificial neural network has highly-parallel processing power, there is self study, self organization ability, can find out the inner link between input and output data according to learning sample, there is Nonlinear Mapping function, be particularly suitable for many nonlinear problems in image processing.In recent years, fuzzy neuron synthesis in theory probabilistic ability and the powerful learning ability of artificial neural network in fuzzy theory analog filtering process, be applied in gradually in the filtering of gray level image.

Summary of the invention

The technical matters solving

Problem to be solved by this invention is that the image filtering method of proposition based on fuzzy neuron system and rim detection, is a kind of image filtering method that can remove spiced salt impulsive noise.

Technical scheme

Technical characterictic of the present invention is, comprises the following steps:

Step 1: by median filter, edge detector and an Adaptive Neuro-fuzzy Inference (ANFIS) combine and form a wave filter, at this wave filter of use, noise image is carried out before filtering, training image of manual construction, use hybrid learning algorithm to train this wave filter, determine the parameter in system;

Step 2: complete when wave filter training, just can carry out filtering to test pattern.

Described step 1 is further comprising the steps:

Steps A: the ANFIS in wave filter has three inputs, an output, using wave filter to carry out, before filtering, need training ANFIS to noise image, determines the parameter in ANFIS.Image of manual construction is as the desired output image of wave filter, adds 30% spiced salt impulsive noise and obtains noise image, as the input picture of ANFIS on this image;

Step B: start with the pixel (this pixel is current operation pixel) in the upper left corner in the training image of inputting, with from top to bottom, mode from left to right, all over getting all pixels in the training image of input on image;

Step C: for current operation pixel, three inputs using median-filtered result, edge detection results and current operation pixel itself as system, can be obtained an actual output of system by three inputs of system;

Step D: choose in the training image of input next pixel as current operation pixel, repeating step C, can obtain the actual output of system of all grey scale pixel values in the training image of input by such mode;

Step e: according to the difference of the actual desired output of exporting and being obtained by the training image of desired output of system of all grey scale pixel values in the training image of input, obtain cost function value;

Step F: in the time that cost function value is less than predefined threshold value, systematic training finishes; Otherwise, use hybrid learning algorithm to be optimized the parameter in system, then repeat above step, carry out next iteration training.

Described step 2 is further comprising the steps:

Steps A: will need the test pattern of filtering as the input picture of wave filter, pixel (this pixel is current operation pixel) with the upper left corner in input picture starts, on image, with from top to bottom, mode from left to right, all over getting all pixels in input picture;

Step B: for current operation pixel, three inputs using median-filtered result, edge detection results and current operation pixel itself as system, can obtain one of system actual output by three inputs of system, this output is the gray-scale value of image pixel after the filtering corresponding with input picture current operation pixel;

Step C: choose in input picture next pixel as current operation pixel, repeating step B, when all pixels in input picture all obtain output through system after, just can obtain an output image, this output image is image after filtering.

Beneficial effect

Impulse noise filter method of the present invention is the spiced salt impulsive noise in filtering image effectively.For the image being polluted by impulsive noise, filtering method proposed by the invention, probabilistic ability and the powerful learning ability of artificial neural network in fuzzy theory analog filtering process are fully utilized, by medium filtering, the information of rim detection and the three aspects: of input picture own combines, in effectively removing image in spiced salt impulsive noise, can retain preferably edge and details in original image, retain visually-perceptible information important in original image, thereby follow-up image is cut apart, the task such as feature extraction and image recognition more can be carried out effectively.

Brief description of the drawings

Fig. 1 is Impulse Noise Filter structural drawing

Fig. 2 is the ANFIS training optimizing process figure for noise filtering

Fig. 3 is the artificial training image for noise filtering

Fig. 4 is the selection of action pane size

Fig. 5 is edge detector structural drawing

Fig. 6 is that four data are extracted the corresponding topological structure of piece

Fig. 7 is the ANFIS training optimizing process figure for rim detection

Fig. 8 is the artificial training image for rim detection

Fig. 9 is Baboon test pattern

Figure 10 be with distinct methods to be subject to varying strength spiced salt impulsive noise pollute Baboon figure denoising after average peak signal to noise ratio (PSNR) comparison curves

Figure 11 is that the Baboon image to being polluted by 40% spiced salt impulsive noise uses the result figure obtaining after distinct methods filtering

Embodiment

Below in conjunction with Fig. 1 to Figure 11, the present invention is described in further detail.

Step 1: by median filter, edge detector and an Adaptive Neuro-fuzzy Inference (ANFIS) combine and form a wave filter, at this wave filter of use, noise image is carried out before filtering, training image of manual construction, use hybrid learning algorithm to train this wave filter, determine the parameter in system;

Concrete steps are as follows:

Steps A: the ANFIS in wave filter has three inputs, an output, using wave filter to carry out, before filtering, need training ANFIS to noise image, determines the parameter in ANFIS.Image of manual construction is as the desired output image of wave filter, adds 30% spiced salt impulsive noise and obtains noise image, as the input picture of ANFIS on this image;

Fig. 1 is proposed Impulse Noise Filter structural drawing.New wave filter is by median filter, and edge detector and an Adaptive Neuro-fuzzy Inference (ANFIS) combine.This ANFIS is one the three single output of input one degree Sugeno fuzzy inference system.Using wave filter to carry out, before filtering, need training ANFIS to test pattern, determine the value of premise parameter and consequent parameter.Fig. 2 is ANFIS training optimizing process figure, and training image wherein can be constructed and be obtained by Artificial.Fig. 3 (a) is the original image of manual construction, is also the training image of the desired output of ANFIS in Fig. 2.This image size is 128 × 128, be made up of 1024 4 × 4 color lumps, 16 pixels in each color lump have identical gray-scale value, and the gray-scale value of different color blocks is different, all over getting all values in 0 to 255, the position of the color lump of different gray-scale values in image is random.Fig. 3 (b) is the training image of the input in Fig. 2, is on Fig. 3 (a) original image, to add the noise image that 30% spiced salt impulsive noise obtains.

Step B: start with the pixel (this pixel is current operation pixel) in the upper left corner in the training image of inputting, with from top to bottom, mode from left to right, all over getting all pixels in the training image of input on image;

Step C: for current operation pixel, three inputs using median-filtered result, edge detection results and current operation pixel itself as system, can be obtained an actual output of system by three inputs of system;

Can obtain corresponding medium filtering image according to input picture.Successively as current operation pixel, centered by this pixel, determine medium filtering window using each pixel in input picture, in window, the intermediate value of all grey scale pixel values is the output after current operation pixel filter.In medium filtering process, need to determine the size of filter window.Known by the performance evaluation to median filter, choosing of intermediate value is subject to the impact of filter window size larger, it is suppressing there is certain contradiction aspect picture noise and protection details two: if the filter window of choosing is less, be conducive to protect some details in image, but can limit the ability of making an uproar of filtering; Otherwise, if the filter window of choosing is larger, can strengthens noise suppression ability, but can weaken the protective capability of details.This contradiction shows particularly evidently in the time that noise in image interference is larger.In the present invention, by simulation results show, according to varying in size of impulsive noise image noise intensity, adopt method in Fig. 4 to determine action pane size, experimental result is best.

Can obtain corresponding edge-detected image according to input picture.In recent years, scholars have proposed the edge detection algorithm of many classics, such as Robert operator, Prewitt operator, Sobel operator, Laplacian of Gaussian (LOG) operator and Canny operator etc.These classical algorithms can extract the edge in image to a certain extent effectively, but the result of rim detection is affected by noise larger.Therefore, adopt these classic algorithm to carry out before Image Edge-Detection, generally need carry out image filtering pre-service.Like this, inevitably improved the complexity of algorithm, and the net result of rim detection affected by filtering larger.To be processed in the present invention is the image that polluted by impulsive noise, and therefore, in the structure of wave filter, these easy algorithms affected by noise are not suitable as the edge detector in system.

In the present invention, we adopt the Image Edge Detector based on Adaptive Neuro-fuzzy Inference shown in Fig. 5.This detecting device comprises four ANFIS and an aftertreatment piece, and each ANFIS is one the four single output of input one degree Sugeno fuzzy inference system.The corresponding data of each ANFIS are extracted piece, and each data are extracted piece and provided four inputs for corresponding ANFIS.For each pixel in input picture, the concrete steps that service data is extracted are as follows:

(1) as shown in Figure 6, with current operation pixel p ₂centered by, obtain 3 × 3 filter windows, obtain grey scale pixel value p according to these four kinds of different topological structures of level, vertical, left diagonal line and right diagonal line ₁, p ₂and p ₃;

(2) still with current operation pixel p ₂centered by, obtain in addition a predefined intermediate value and extract action pane, by the gray-scale value of all pixels in this window, obtain intermediate value m, intermediate value is extracted determining as shown in Figure 4 of action pane size;

(3) make four of each ANFIS to input x ₁, x ₂, x ₃and x ₄be respectively:

$\left\{\begin{array}{c}{x}_1=p_1-m\\ x_2=p_2-m\\ x_3=p_3-m\\ x_4=m\end{array}\right.,---\left(1\right)$

In Fig. 5, the inner structure of four ANFIS is all identical, is all one the four single output of input one degree Sugeno fuzzy inference system.For the current operation pixel in input picture, each data are extracted piece and are provided four input x for corresponding ANFIS ₁, x ₂, x ₃and x ₄.For each input, define respectively three broad sense bell subordinate functions, each ANFIS comprises 81 (3 ⁴) rule.Each ANFIS can obtain an output, and four ANFIS can obtain four outputs, are designated as Y _k(k=1,2,3,4), these four outputs are the input of aftertreatment piece.Aftertreatment piece is averaged to these four inputs by formula (2), is designated as Y _a.Again by formula (3), by Y _awith a threshold, try to achieve the final output Y of edge detector _f.Threshold value is wherein the intermediate value between minimum value and the maximal value of grey scale pixel value, and herein, for 8 gray level images, this value is set as 128.Final output Y _fvalue be that 0 expression current operation pixel is edge pixel, be shown as black, Y _fvalue be that 255 expression current operation pixels are not edge pixels, be shown as white.

$Y_A=\frac{1}{4}\underset{k=1}{\overset{4}{\mathrm{\Σ}}}{Y}_k---\left(2\right)$

$Y_F=\left\{\begin{array}{cc}255,& \mathrm{if}{Y}_A>128\\ 0,& \mathrm{if}{Y}_A\≤128\end{array}\right.---\left(3\right)$

At edge detector shown in use Fig. 5, input picture is carried out before rim detection, each ANFIS needs to train separately.Fig. 7 is single ANFIS training optimizing process figure, and training image wherein can be constructed and be obtained by Artificial, and each ANFIS adopts identical training image.Fig. 8 (a) is the original training image of each ANFIS, and this image is identical with Fig. 3 (a).Fig. 8 (b) is the training image of the input in Fig. 7, and this image is identical with Fig. 3 (b).Fig. 8 (c) is the edge flag image obtaining according to Fig. 8 (a) original image, the namely training image of desired output in Fig. 7, in figure, grey scale pixel value is that 0 expression current pixel is edge pixel, be shown as black, grey scale pixel value is that 255 expression current pixels are not edge pixels, is shown as white.

Premise parameter in each ANFIS and consequent parameter need by training definite its value, and the optimized algorithm using when system is trained is hybrid learning algorithm.When four ANFIS train completely, just can form edge detector (as shown in Figure 5) together with an aftertreatment piece, input picture is carried out to rim detection, obtain edge flag image.

Obtain after corresponding medium filtering image and edge-detected image according to input picture, just can with input picture itself, as three inputs of system shown in Figure 2, can obtain one of system actual output by three inputs of system.For the current operation pixel in input picture, three inputs of establishing this system are respectively x ₁, x ₂, x ₃, wherein x ₁for the median-filtered result of current operation pixel, x ₂for the edge detection results of current operation pixel, x ₃for current operation pixel itself.For each input, define respectively three broad sense bell subordinate functions, this system comprises 27 (3 altogether ³) rule, its Fuzzy Rule Sets is as follows:

Rule 1:if (x ₁isM ₁₁) and (x ₂isM ₂₁) and (x ₃isM ₃₁)

then?y ₁＝d ₁₁x ₁+d ₁₂x ₂+d ₁₃x ₃+d ₁₄

Rule 2:if (x ₁isM ₁₁) and (x ₂isM ₂₁) and (x ₃is M ₃₂)

then?y ₂＝d ₂₁x ₁+d ₂₂x ₂+d ₂₃x ₃+d ₂₄

Rule 3:if (x ₁isM ₁₁) and (x ₂isM ₂₁) and (x ₃isM ₃₃)

then?y ₃＝d ₃₁x ₁+d ₃₂x ₂+d ₃₃x ₃+d ₃₄

Rule 4:if (x ₁isM ₁₁) and (x ₂is M ₂₂) and (x ₃isM ₃₁)

then?y ₄＝d ₄₁x ₁+d ₄₂x ₂+d ₄₃x ₃+d ₄₄

|

|

|

Rule 27:if (x ₁isM ₁₃) and (x ₂is M ₂₃) and (x ₃is M ₃₃)

then?y ₂₇＝d _27，1x ₁+d _27，2x ₂+d _27，3x ₃+d _27，4

Wherein M _ijrepresent j subordinate function of i input, d _klfor true value parameter, y _kfor the output that system obtains according to k rule, i=1,2,3, j=1,2,3, k=1,2 ..., 27, l=1,2,3,4.For input x _i, the broad sense bell subordinate function of definition is:

$M_{\mathrm{ij}}\left(x_i\right)=\frac{1}{1+{\left|\frac{x_i-c_{\mathrm{ij}}}{a_{\mathrm{ij}}}\right|}^{{2b}_{\mathrm{ij}}}},---\left(4\right)$

The output Y of this system equals each y _kweighted mean:

$Y=\frac{\underset{k=1}{\overset{27}{\mathrm{\Σ}}}{w}_k{y}_k}{\underset{k=1}{\overset{27}{\mathrm{\Σ}}}{w}_k}---\left(5\right)$

Weighting coefficient w in formula _kcomprise that k acting rules, in the obtained all true value of input, represent the excitation density of k rule, w _kcomputing formula as follows:

$\begin{array}{c}{w}_1=M_{11}\left(x_1\right)\×M_{21}\left(x_2\right)\×M_{31}\left(x_3\right)\\ w_2=M_{11}\left(x_1\right)\×M_{21}\left(x_2\right)\×M_{32}\left(x_3\right)\\ w_3=M_{11}\left(x_1\right)\×M_{21}\left(x_2\right)\×M_{33}\left(x_3\right)\\ w_4=M_{11}\left(x_1\right)\×M_{22}\left(x_2\right)\×M_{31}\left(x_3\right)\\ .\\ .\\ .\\ w_{27}=M_{13}\left(x_1\right)\×M_{23}\left(x_2\right)\×M_{33}\left(x_3\right)\end{array}---\left(6\right)$

Step D: choose in the training image of input next pixel as current operation pixel, repeating step C, can obtain the actual output of system of all grey scale pixel values in the training image of input by such mode;

Step e: according to the difference of the actual desired output of exporting and being obtained by the training image of desired output of system of all grey scale pixel values in the training image of input, obtain cost function value;

Step F: in the time that cost function value is less than predefined threshold value, systematic training finishes; Otherwise, use hybrid learning algorithm to be optimized the parameter in system, then repeat above step, carry out next iteration training.

To the parameter in ANFIS, adopt hybrid learning algorithm to be optimized, i.e. premise parameter a _ij, b _ijand c _ijbe optimized consequent parameter d by gradient method _klbe optimized with linear least square, the learning strategy of employing is off-line (in batches) learning method.

Suppose to have N group training data, for each group input x _t=(x ₁, x ₂, x ₃) ^t, actual output and the desired output of system are respectively Y _tand Yd _t, t=1,2 ..., N, T representing matrix transposition, definition cost function:

$E=\underset{t=1}{\overset{N}{\mathrm{\Σ}}}{E}_t=\underset{t=1}{\overset{N}{\mathrm{\Σ}}}\left(\frac{1}{2}{({\mathrm{Yd}}_t-Y_t)}^2\right)---\left(7\right)$

For premise parameter a _ij, b _ij, c _ij, have according to gradient method:

$a_{\mathrm{ij}}(n+1)=a_{\mathrm{ij}}\left(n\right)-\mathrm{\α}\frac{\∂E}{\∂a_{\mathrm{ij}}}---\left(8\right)$

$b_{\mathrm{ij}}(n+1)=b_{\mathrm{ij}}\left(n\right)-\mathrm{\α}\frac{\∂E}{\∂b_{\mathrm{ij}}}---\left(9\right)$

$c_{\mathrm{ij}}(n+1)=c_{\mathrm{ij}}\left(n\right)-\mathrm{\α}\frac{\∂E}{\∂c_{\mathrm{ij}}}---\left(10\right)$

Wherein n is iterations, and α is learning rate, i=1,2,3, j=1,2,3.When training, can organize training data according to N and determine premise parameter a _ij, b _ijand c _ijinitial value.

For consequent parameter d _klwe are first according to the value of premise parameter, formula (5) and N group training data (comprising input and desired output), adopt linear least square method to determine its initial value, then can be obtained the actual output of system of N group training data by formula (5), then upgrade premise parameter a according to the cost function of formula (7) and formula (8), formula (9) and formula (10) _ij, b _ij, c _ij, and after each renewal premise parameter, adopt linear least square method to upgrade consequent parameter d _kl.Iteration is gone down so always, until the cost function value of formula (7) is less than predefined threshold value or iterations reaches predefined in limited time upper, systematic training finishes.

Step 2: complete when wave filter training, just can carry out filtering to test pattern.

Concrete steps are as follows:

Steps A: will need the test pattern of filtering as the input picture of wave filter, pixel (this pixel is current operation pixel) with the upper left corner in input picture starts, on image, with from top to bottom, mode from left to right, all over getting all pixels in input picture;

For the filtering performance of test the inventive method, it is compared with traditional filtering method, Baboon figure (image size is 256 × 256) is carried out to test analysis, as shown in Figure 9.Baboon figure is added to the test pattern of 3%～80% spiced salt impulsive noise as input.

Step B: for current operation pixel, three inputs using median-filtered result, edge detection results and current operation pixel itself as system, can obtain one of system actual output by three inputs of system, this output is the gray-scale value of image pixel after the filtering corresponding with input picture current operation pixel;

First obtain corresponding medium filtering image and edge-detected image according to the test pattern of input, then with the test pattern itself of inputting, as three inputs of system shown in Figure 1, can be obtained an actual output of system by three inputs of system.

Step C: choose in input picture next pixel as current operation pixel, repeating step B, when all pixels in input picture all obtain output through system after, just can obtain an output image, this output image is image after filtering.

For the filter effect of the inventive method and classic method is compared, select following filtering method: SMF method, EDMF method, MSMF method, SDROMF method, PSMF method and FF method.In experiment, we adopt the performance of Y-PSNR PSNR evaluation criterion quantitative evaluation wave filter.PSNR value shows that more greatly filter filtering effect is better.PSNR is defined as:

$\mathrm{PSNR}={10\log }_{10}\left(\frac{{255}^2}{\frac{1}{\mathrm{MN}}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{(O(i,j)-R(i,j))}^2}\right)---\left(11\right)$

Wherein O (i, j) and R (i, j) are respectively the gray-scale value of image slices vegetarian refreshments after coordinate points (i, j) is located original noise-free picture and filtering, and image size is M × N.

Figure 10 has provided SMF method, EDMF method, MSMF method, SDROMF method, PSMF method, FF method and the inventive method to be subject to 3%～80% intensity spiced salt impulsive noise pollute Baboon image denoising after average peak signal to noise ratio (PSNR) comparison curves.From figure, can find out significantly, to the image that polluted by varying strength spiced salt impulsive noise, the PSNR that the inventive method obtains is large, and this has illustrated that the more traditional filtering method of the filter effect of the inventive method is for well.

For the filtering performance of evaluating a wave filter, except above-mentioned quantitative evaluation, also need to carry out qualitative evaluation.Subjective qualitative evaluation is exactly the filter effect from visually observing various wave filters.What Figure 11 showed is the filter effect comparison diagram of Baboon figure.Figure 11 (a) is original image (not by image polluted by noise) Baboon figure, Figure 11 (b) is the spiced salt impulsive noise figure containing 40%, Figure 11 (c)～(h) is respectively the result images that each traditional filtering method is exported, the result images that Figure 11 (i) exports for the inventive method.In filter effect figure by the inventive method and traditional filtering method, can find out, in the output image of conventional traditional filtering method (SMF, EDMF, MSMF, SDROMF and FF), have and significantly there is no removed noise spot, particularly, in the output image of SMF and MSMF method, do not have removed noise spot a lot.And in the output image of the inventive method, noise remove must be cleaner, this explanation, the ability that the inventive method is removed impulsive noise is strong compared with traditional filtering method.In addition, compared with the Output rusults of traditional filtering method, the inventive method more can retain details or the edge of original image, and this can obviously find out from the around eyes of Baboon and beard position.Figure 11 (g) is the filtering result figure of PSMF method, although there is no obvious noise spot in figure, its newer filtering method of ability that retains original image detail is for poor, and after its filtering, image can obviously be found out blur effect.

Claims (3)

1. the image filtering method based on fuzzy neuron system and rim detection, is characterized in that, comprises the following steps:

Step 1: by median filter, edge detector and an Adaptive Neuro-fuzzy Inference (ANFIS) combine and form a wave filter, at this wave filter of use, noise image is carried out before filtering, training image of manual construction, use hybrid learning algorithm to train this wave filter, determine the parameter in system;

Step 2: complete when wave filter training, just can carry out filtering to test pattern.

2. the image filtering method based on rim detection according to claim 1, is characterized in that, step 1 is further comprising the steps:

Steps A: the ANFIS in wave filter has three inputs, an output, using wave filter to carry out, before filtering, need training ANFIS to noise image, determines the parameter in ANFIS.Image of manual construction is as the desired output image of wave filter, adds 30% spiced salt impulsive noise and obtains noise image, as the input picture of ANFIS on this image;

Step B: start with the pixel (this pixel is current operation pixel) in the upper left corner in the training image of inputting, with from top to bottom, mode from left to right, all over getting all pixels in the training image of input on image;

Step C: for current operation pixel, three inputs using median-filtered result, edge detection results and current operation pixel itself as system, can be obtained an actual output of system by three inputs of system;

Step D: choose in the training image of input next pixel as current operation pixel, repeating step C, can obtain the actual output of system of all grey scale pixel values in the training image of input by such mode;

Step e: according to the difference of the actual desired output of exporting and being obtained by the training image of desired output of system of all grey scale pixel values in the training image of input, obtain cost function value;

Step F: in the time that cost function value is less than predefined threshold value, systematic training finishes; Otherwise, use hybrid learning algorithm to be optimized the parameter in system, then repeat above step, carry out next iteration training.

3. the image filtering method based on rim detection according to claim 1, is characterized in that, step 2 is further comprising the steps:

Steps A: will need the test pattern of filtering as the input picture of wave filter, pixel (this pixel is current operation pixel) with the upper left corner in input picture starts, on image, with from top to bottom, mode from left to right, all over getting all pixels in input picture;

Step B: for current operation pixel, three inputs using median-filtered result, edge detection results and current operation pixel itself as system, can obtain one of system actual output by three inputs of system, this output is the gray-scale value of image pixel after the filtering corresponding with input picture current operation pixel;

Step C: choose in input picture next pixel as current operation pixel, repeating step B, when all pixels in input picture all obtain output through system after, just can obtain an output image, this output image is image after filtering.