Method for retrieving original intact characteristics of heavily polluted images and its image processing

Abstract

A method for retrieving original intact characteristics of images heavily polluted by Gaussian impulse noises is based on the fuzzy automata theory retrieves the characteristics of images heavily corrupted by Gaussian impulse noises. The retrieved characteristics are represented with fuzzy sets, each of which is also called a fuzzy partition of the image histogram. The membership functions are used for noise filtering, multi-level thresholding, segmentation, and edge detection. This invention further utilizes the result images of filtering and thresholding to perform the second phase processing. After two-phase filtering, the signal-to-noise ratio SNR and the peak signal to noise ratio PSNR are increased up to 10 db for images polluted by noises higher than 50%. Having very low complexity of processing time and memory space, the kernel operation of the algorithm can be implemented with super-scalar super-pipelined 16-bit floating-point processing unit and achieve 24-bit true color results.

Description

FIELD OF THE INVENTION

Based on fuzzy automata theory, this invention describes a novel method that retrieves the original intact characteristics of images heavily polluted by Gaussian impulse noises. The characteristics are represented with fuzzy sets whose membership functions are used for noise removal, multi-level thresholding, segmentation, edge detection, and other image analysis operations. The image results after the first phase (Phase I) processing including noise removal and multi-level thresholding are used for further removal of corrupted spots and for further analyses, and these further operations are called the second phase (phase II) image processing. After the second phase processing, the result image becomes much clear according to not only the feeling of subjective eyesight but also the objective and evident increasing of SNR and PSNR values. At the same time, we obtain the images of multi-level thresholding, segmentation, and edge detection. These result images then contain enough useful information for analyses in fields such as medical diagnosis, image compression, image recognization, . . . , etc.

BACKGROUND OF THE INVENTION

Image is the most effective and most direct way to communicate. Nowadays, image transportation is widely and commonly applied in medicine, security, space, and many other sciences. However, due to much kind of possible defects of transportation media, communication channel, imaging facility, and/or failure of compression and storage, a clean image could be heavily corrupted by noises. Due to loss of much important information, the corrupted image is useless and loses its value in any industrial application. Before presenting the image to a user, we need to recover its information for analysis purposes such as multi-level thresholding, edge detection, and image segmentation. In image processing area, there are many filtering methods for noise removal such as low pass filters, median filter, weighted median filters (WM, Browrigg, 1984), center weighted median filter (CWM, Ko and Lee, 1991), and adaptive weighted fuzzy mean filters (AWFM, Kuo and Lee, 2000). Most of these methods apply only one-pass noise removal. After single pass noise removal, there are still large amount of corrupted spots in the image. For example, in sorting based filters, when the noise ratio is higher than 30%, being median gray level in the sample window, a noisy pixel is regarded as intact and will be filled back to the center of the sample window. Therefore, the median operation could further destroy the image by using many noisy pixels as operands. For the case in AWFM filters, the gray levels are partitioned into fixed three classes—Dark, Median, and Bright. The number of classes and distribution of gray levels in each class are fixed and are not adaptively obtained according to the characteristics of the corresponding image under processing. Thus, the result of WFM filters will be unacceptable when the gray levels of an image cannot be partitioned into three classes. This invention adaptively and automatically retrieves characteristics of each image under processing and clean up heavy noises by two-phase processing scheme. The second phase processing utilizes the result of the first phase to produces stable and the best result among the state-of-art filters.

In the field of image analysis, the primary processing operations are thresholding, segmentation, and edge detection. This invention not only filters out image noises but at the same time also performs the operations in these analyses. For highly polluted images such as those having noise ratio higher than 30%, there have not been methods to pursue thresholding, segmentation, and edge detection at low costs in both time and space complexity. Therefore, the invention does make a great advance in image analysis.

SUMMARY OF THE INVENTION

The invention primarily targets a noise filtering and analyses of thresholding, segmentation, and edge detection for heavily corrupted images by using a two-phase processing scheme. In addition to recovering clean image most close to the original and performing essential analyses, the processing time and the memory space costs are dramatically reduced compared with state-of-art methods.

The primary feature of this invention is to provide a two-phase image-processing algorithm. In the first phase the algorithm produces histogram of the corrupted image and according to the histogram automatically generate fuzzy partitions and their corresponding membership functions represented by LR parameters. The LR parameters representing an LR fuzzy set (α, m, β)_LR are the left spread α, the mean m, and the right spread β. With the automatically generated fuzzy partitions and their corresponding membership functions, the algorithm filters the corrupted image and produces the Phase-I multi-level thresholded image. In the second phase processing, the algorithm re-samples the image results of the first phase. In the sample window, the algorithm compares the center pixel with its surrounding ones to determine whether the center pixel is corrupted and takes the median gray level of the intact pixels in the sample window as a new center pixel.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1. Image sample window and its moving behavior when scanning pixels.

FIG. 2. The diagram showing the concept of using fuzzy inference for image processing.

FIG. 3. The flowchart of Phase I image filtering and processing.

FIG. 4. The flowchart of Phase II image filtering and processing.

FIG. 5. The diagram of the two-phase image processing of this invention.

FIG. 6. shows the processing of the present invention, wherein

(a) Baboon original image

(b) Baboon image with additive 70% Gaussian impulse noise of mean length 150.

(c) Baboon image filtered with the MEDIAN Filters.

(d) Baboon image filtered with the WFM Filter.

(e) Baboon image filtered with the Phase I AMFG Filter.

(f) Baboon image filtered with the Phase II AMFG Filter.

FIG. 7. shows a plurality drawings of the present invention, where

(a) Baboon image thresholded after filtering of WFM Filter.

(b) Baboon image thresholded after filtering of Phase I AMFG Filter.

(c) Baboon image thresholded after filtering of Phase II AMFG Filter.

(d) Edge detection of the Baboon image after WFM Filter

(e) Edge detection of the Baboon image after Phase I AMFG Filter.

(f) Edge detection of the Baboon image after Phase II AMFG Filter.

FIG. 8. shows a plurality drawings of the present invention, where

(a) Lenna original image

(b) Lenna image with additive 70% Gaussian impulse noise of mean length 150.

(c) Lenna image filtered with the MEDIAN Filters.

(d) Lenna image filtered with the WFM Filter.

(e) Lenna image filtered with the Phase I AMFG Filter.

(f) Lenna image filtered with the Phase II AMFG Filter.

FIG. 9. shows a plurality drawings of the present invention, where

(a) Lenna image thresholded after filtering of WFM Filter.

(b) Lenna image thresholded after filtering of Phase I AMFG Filter.

(c) Lenna image thresholded after filtering of Phase II AMFG Filter.

(d) Edge detection of the Lenna image after WFM Filter

(e) Edge detection of the Lenna image after Phase I AMFG Filter.

(f) Edge detection of the Lenna image after Phase II AMFG Filter.

FIG. 10. shows a plurality drawings of the present invention, where

(a) Peppers original image

(b) Peppers image with additive 70% Gaussian impulse noise of mean length 150.

(c) Peppers image filtered with the MEDIAN Filters.

(d) Peppers image filtered with the WFM Filter.

(e) Peppers image filtered with the Phase I AMFG Filter.

(f) Peppers image filtered with the Phase II AMFG Filter.

FIG. 11. shows a plurality drawings of the present invention, where

(a) Peppers image thresholded after filtering of WFM Filter.

(b) Peppers image thresholded after filtering of Phase I AMFG Filter.

(c) Peppers image thresholded after filtering of Phase II AMFG Filter.

(d) Edge detection of the Peppers image after WFM Filter

(e) Edge detection of the Peppers image after Phase I AMFG Filter.

(f) Edge detection of the Peppers image after Phase II AMFG Filter.

FIG. 12. shows a plurality drawings of the present invention, where

(a) Boat original image

(b) Boat image with additive 70% Gaussian impulse noise of mean length 150.

(c) Boat image filtered with the MEDIAN Filters.

(d) Boat image filtered with the WFM Filter.

(e) Boat image filtered with the Phase I AMFG Filter.

(f) Boat image filtered with the Phase II AMFG Filter.

FIG. 13. shows a plurality drawings of the present invention, where

(a) Boat image thresholded after filtering of WFM Filter.

(b) Boat image thresholded after filtering of Phase I AMFG Filter.

(c) Boat image thresholded after filtering of Phase II AMFG Filter.

(d) Edge detection of the Boat image after WFM Filter

(e) Edge detection of the Boat image after Phase I AMFG Filter.

(f) Edge detection of the Boat image after Phase II AMFG Filter.

BRIEF DESCRIPTION OF THE INVENTION

To achieve the analysis and filtering purposes and performance mentioned above, we depict the proposed techniques by using the following illustrated examples with figures.

In the field of image processing, the filtering is similar to that in communication. Both of them in turns sample the input, perform operation, filtering or processing, and then output the result. Commonly used sampling mechanism in image processing is to use a two-dimensional sample window to sample the image pixels in one dimension first. Each time the sample window moves one pixel along the dimension. As FIG. 1, the sample window moves horizontally first to scan the image pixels. When the sample window reaches the end of the dimension, it moves vertically down one pixel along the other dimension and again starts the horizontal scanning. The iteration of this manner repeats until all pixels are scanned.

This invention uses fuzzy inference for image processing as FIG. 2. The design specification and the algorithm determine the size of the sample window. The most commonly used sample window size is 3-by-3 (3×3). Adopting odd-by-odd sample window usually produce better results and among odd-by-odd sample windows, the 3-by-3 one is usually superior in filtering. This is because an image is a continuous two-dimensional signal with spatial locality. Thus, in a noise-free 3-by-3 sample window, the center pixel will have similar gray-level compared with surrounding pixels. The fuzzy inference system processes the pixels in the sample window and produces a new gray level to replace one of the pixels in the sample window. In the following description, we constantly replace the center pixel in the sample window.

This invention proposes the analyses of heavily corrupted images using fuzzy partitions in two-phases. We depict these two phases as follows:

1. Phase I

In the first phase filtering, according to the characteristics of the image under processing, the method automatically retrieves the parameters α, m, and β of the membership functions of the fuzzy partitions and then uses these membership functions to filter noises. The proposed filter is much superior to the sorting based filters and is also better than the weighted fuzzy mean filters, which adopt fixed membership functions. In addition, fuzzy partitions of the histogram also provide multi-level thresholds such that all gray levels in the same partition are assigned with a new representative gray level. That is, in addition to filtering, the Phase I multi-level thresholding is obtained at the same time.

Please refer to FIG. 3 and the processing steps is as follows:

Step 1. Generate the Histogram and Partition the Histogram into Coarse Sections.

Read the N-by-N (N×N) image pixels into array XA(i, j). Accumulate the number of appearance of each gray level x as the histogram hs(x). Drop the statistics of x if x is outside the interval [5, L−5], where L is the largest gray level. In this way most impulse noisy pixels will not taken into account for membership function parameters calculations. The first x with nonzero statistic is the start point hs_s, of the histogram and the final x with nonzero statistic is the end point hs_f of the histogram.

To find out the fuzzy partitions of hs(x), we perform down-sampling of the histogram as equations (1) and (2), where v is the sampling reference value, c_n is the number of cut section, c_p is the number of gray levels in each section.

$\begin{array}{cc}{c}_n=\left(\left(\nu \times 2+1\right)+2\right)\times 2& \left(1\right)\\ c_p=\lfloor \frac{{\mathrm{hs}}_f-{\mathrm{hs}}_s}{c_n}\rfloor & \left(2\right)\end{array}$

In this way we down sample the histogram into C_n levels.

Step 2. Find Out Local Minimum (Valley Point) and Local Maximum (Peak Point) of The Original Histogram.

Calculate the starting point c_hs(i) of each section i in hs(x) and the average value r_hs(i) as (3) and (4).

$\begin{array}{cc}{c}_{\mathrm{hs}}\left(i\right)=\lfloor \left(\frac{{\mathrm{hs}}_f-{\mathrm{hs}}_s}{c_n}\right)\times i\rfloor & \left(3\right)\\ r_{\mathrm{hs}}\left(i\right)=\frac{\sum _{x=c_p\left(i-1\right)+1}^{c_p\times i}\mathrm{hs}\left(x\right)}{c_p}& \left(4\right)\end{array}$

Calculate each difference r_o(i) between every two consecutive sections as (5).r_o(i)=r_hs(i+1)−r_hs(i)  (5)

• • Determine whether c_hs(i) is a coarse valley point c_m(i) as (6). Between two initial valley points c_m(i) and c_m(i−1) find the peak point c_M(i) as (7).

$\begin{array}{cc}\left\{c_m\left(k\right)\text{|}k\in N\right\}=\left\{c_{\mathrm{hs}}\left(i\right)\text{|}{r}_o\left(i\right)>0\mathrm{and}{r}_o\left(i-1\right)<0,i\in N-\left\{1\right\}\right\}& \left(6\right)\\ c_M\left(i\right)=\underset{x\in \left[c_m\left(i-1\right),c_m\left(i\right)-1\right)]}{\max }\mathrm{hs}\left(x\right)& \left(7\right)\end{array}$

Between two peak points c_M(i) and c_M(i−1), find the accurate valley point c_m(i) as (8).

$\begin{array}{cc}{c}_m\left(i\right)=\underset{x\in \left[c_M\left(i-1\right),c_M\left(i\right)-1\right]}{\min }\mathrm{hs}\left(x\right)& \left(8\right)\end{array}$

Step 3. Calculate Parameters α_i, m_i, and β_i of each Fuzzy Partition i as (9), (10), and (11) and Obtain the Corresponding Membership Functions.

The left spread parameter:α_i=c_M(i)−c_m(i)  (9)

The mean parameter:m_i=c_M(i)  (10)

The right spread:β_i=c_m(i+1)−c_M(i)  (11)

We use the LR fuzzy set (α_i, m_i, β_i)_LR to represent fuzzy partition i. Use sample window YA of size s_h×s_w to sample XA. The (h, w)-th sample window is as (12).YA(i,j)=XA(i,j)iε[h,h+(s_h−1)]jε[w,w+(s_w−1)]  (12)

Substitute YA(i, j) into membership function of the k-th fuzzy partition (α_k, m_k, β_k)_LR to obtain pixel YA(i,j)'s k-th membership grade μ_k(i,j) as (13).

$\begin{array}{cc}{\mu }_k\left(i,j\right)=\mathrm{LR}\left(\frac{m_k\ominus \mathrm{YA}\left(i,j\right)}{{\alpha }_k}+\frac{\mathrm{YA}\left(i,j\right)\ominus m_k}{{\beta }_k}\right),& \left(13\right)\end{array}$ where a⊖b=max(a−b, 0) and LR(x) is the reference function such as LR(x)=1⊖x.

Step 4. Calculate the Maximum Likelihood Estimator and the Weighted Mean.

Calculate the maximum likelihood estimator AV_hw of pixels in sample window YA as (14). Summation of the multiplication of pixel YA(i, j) and weight μ_k(i, j) to obtain the weighted value w_k,ux as (15). Divide w_k,ux by w_k,u as (16) and round the division to obtain the weighted mean w_k,am as (17).

$\begin{array}{cc}{\mathrm{AV}}_{\mathrm{hw}}=\frac{\sum _{i=1}^{s_w}\sum _{j=1}^{s_h}\mathrm{YA}\left(i,j\right)}{s_h\times s_w}& \left(14\right)\\ w_{k,\mathrm{ux}}=\sum _{i=1}^{s_w}\sum _{j=1}^{s_h}\mathrm{YA}\left(i,j\right)\times {\mu }_k\left(i,j\right)& \left(15\right)\\ w_{k,u}=\sum _{i=1}^{s_w}\sum _{j=1}^{s_h}{\mu }_k\left(i,j\right)& \left(16\right)\\ w_{k,\mathrm{am}}=\lfloor \left(\left(w_{k,\mathrm{ux}}/w_{k,u}\right)+0.5\right)\rfloor & \left(17\right)\end{array}$

Step 5. Output the Filtering Result and Obtain the Thresholding.

Calculate the difference between the k-th weighted mean w_k,am and the maximum likelihood estimator AV_hw. Among w_k,am's, find the one w_k,gm closest to AV_hw as (18). Assign OK(h, w) with w_k,gm. Repeat Step 4 and Step 5 until all N×N pixels in image XA are sampled and processed. That is when h=N and w=N the matrix OK is the resultant image of the first phase filtering.

$\begin{array}{cc}{w}_{g,\mathrm{am}}=\underset{w_{k,\mathrm{am}}}{\arg }\left\{\underset{k}{\min }{\mathrm{AV}}_{\mathrm{hw}}-w_{k,\mathrm{am}}\right\}& \left(18\right)\end{array}$

Perform thresholding over OK(h, w) and obtain the thresholded image TH(h, w) as (19).(∀iεN−{1},c_m(i)≦OK(h,w)<c_m(i+1))[TH(h,w) =c_M(i)]  (19)Matrix TH is the Phase I thresholded image.2. Phase II

In the second phase processing, since the thresholded image TH from the first phase has much reduced gray level complexity, TH is used for judgment of whether a center pixel of a sample window is corrupted. When some surrounding pixels in the sample window have different gray level from the center pixel's, there are two expectances—the center pixel is either part of an edge segment or corrupted. In a sample window, the center pixel is compared with its surrounding ones. When the amount of pixels of the same gray level is less than a reference value Kt, it is regarded as part of a corrupted spot. The phase II algorithm replaces this pixel's gray level with the median of the non-corrupted ones in the sample window. Please refer to FIG. 4 and the detail steps is described as follows:

Step 1. Read the Thresholded Image in the First Phase into a Matrix TH of size N×N.

Step 2. Sample Matrix TH with a Sample Window xd of Size n×n, For Matrix OK, we Also Use an n×n Sample Window xo for Pixel Scanning.

• • The (I, J)-th sample window is as (20).xd(i,j)=TH(i,j)xo(i,j)=OK(i,j)iε[I,I+(n−1)]jε[J,J+(n−1)]  (20)

Step 3. Set the Reference Value Kt as Accumulating Threshold for Corruption Determination. Compare the Center Pixel of the Sample Window xd with the Surrounding Pixels One by One.

• • Let K has initial value 0.
  • For each comparison, if the gray level is the same, increase K.
  • Otherwise, K remains unchanged.

Step 4. Compare the Value of K with Kt.

• • If K>Kt, the center pixel remain unchanged and we go back to Step 2 to move the two sample windows to the next position. Otherwise, the center pixel is regarded corrupted and proceed to Step 5 and Step 6.

Step 5. Find the Median Value of Uncorrupted Pixels in Sample Window xo.

• • Since now the center pixel is regarded as corrupted, in the sample window xd, the pixels different from the center xd(I+┌n/2┐,J+┌n/2┌) are regarded uncorrupted and only their corresponding pixels in xo can join the median calculation. We replace the corrupted center pixel with the median gray level of xo.

Step 6. Replace the Center Pixel xo(I+┌n/2┐,J+┌n/2┌) with the Median Obtained in Step 5.

• • When the sample window moves to the final position, that is, I=N and J=N. The matrix OK is updated and the second phase filtering is end.

Step 7. Obtain the New Thresholded Image TH from OK as Previously Described in (19).

Step 8. Detect the Edge and Obtain the Edge Image Ed.

• • As (21), when the sample window xd of matrix TH has pixels of more than two gray levels, the center pixel is regarded a part of an edge.(∀I,J≦N)(∀i,j≦n)(xd(I+┌n/2┐,J+┌n/2┐≠xd(I+i,J+j)) [ED(I,J)]=255  (21)

The above two-phase filtering is shown as FIG. 5. We call the proposed filter the automatic membership function generation (AMFG) filter. The AMFG filter adaptively generate fuzzy partitions according to the characteristics of the image under processing and hence has better noise removal ability than fuzzy mean filters even though they both use fuzzy weighted mean for filtering. Moreover, in Phase II, the AMFG Filter further removes corrupted spots left in the resultant image of Phase I. The removal utilizes the two results including filtered and thresholded images from Phase I. The operations use only non-corrupted pixels and thus the probability that noises participate in the calculation is very low. From the experiments as shown in FIG. 6 to FIG. 13 and according to the following tables comparing SNR and PSNR values, AMFG filter is superior to both WFM and sorting-based filters. The distribution of Gaussian impulse noise in these tables is represented as N(μ, σ) where μ is the mean and σ² is the variance of the distribution. That is when we regard the Gaussian impulse noises as two sources—Gaussian and impulse noises, the impulse noise length is ±μ and the white Gaussian noise is distributed with zero mean and variance σ².

TABLE 1
SNR comparisons of Baboon image with Gaussian impulse noise N(100, 1)
Baboon100	10%	20%	30%	40%	50%	60%	70%	80%	90%	100%
Source	12.975	10.0727	8.2699	7.0757	6.0696	5.2769	4.6151	4.0348	3.5264	3.067
Image
Median	18.1087	17.4394	16.02	14.2473	12.0812	10.0857	8.1744	6.5362	5.1474	3.9777
Filtered
WFM	16.9384	16.1784	15.2772	14.3058	13.4519	12.3866	10.9994	8.3095	5.6781	3.1844
Filtered
Phase I	16.441	15.6664	14.6333	14.4744	13.6627	12.6289	11.1135	8.7646	5.8662	4.4397
Filtered
Phase II	16.5153	15.8941	15.1072	15.6946	15.3717	14.7763	13.6184	10.7106	6.3366	4.5263
Filtered
Phase III	16.3854	15.8138	15.0586	15.6915	15.4878	15.0931	14.1858	11.4022	6.5224	4.5538
Filtered

TABLE 2
SNR comparisons of Baboon image with Gaussian impulse noise N(150, 1)
Baboon150	10%	20%	30%	40%	50%	60%	70%	80%	90%	100%
Source	10.8244	7.768	5.9896	4.7329	3.7774	2.9929	2.2834	1.7313	1.2292	0.7659
Image
Median	18.0166	16.9213	15.0657	12.5092	9.919	7.5337	5.3065	3.6388	2.2452	1.0347
Filtered
WFM	17.5799	16.9728	16.0106	14.6819	13.0505	11.3619	8.9688	6.4711	3.4843	0.3572
Filtered
Phase I	16.533	15.9187	14.8974	14.0902	12.7487	11.157	9.4444	6.4916	3.5728	3.078
Filtered
Phase II	16.6718	16.2471	15.6855	15.6911	15.0578	14.2947	12.6084	8.9438	4.3779	3.2899
Filtered
Phase III	16.5768	16.2	15.657	15.7529	15.3631	14.832	13.5121	9.9552	4.7145	3.3234
Filtered

TABLE 3
SNR comparisons of Lenna image with Gaussian impulse noise N(100, 1)
Lenna100	10%	20%	30%	40%	50%	60%	70%	80%	90%	100%
Source	12.7738	9.8454	8.1084	6.8865	5.9126	5.1209	4.4449	3.8921	3.3693	2.9216
Image
Median	23.0554	21.0961	18.553	15.3751	12.5733	9.9956	7.8829	6.1276	4.6346	3.4728
Filtered
WFM	20.1453	18.8165	17.5187	16.0191	14.7736	13.1493	11.4134	9.0531	6.0836	2.9105
Filtered
Phase I	20.7353	19.751	18.5224	16.9775	15.5831	13.7298	11.7393	8.0751	6.2016	0.8327
Filtered
Phase II	20.2974	19.7845	19.341	18.8275	17.9098	16.7958	14.8107	9.8009	7.0708	0.9592
Filtered
Phase III	20.0214	19.4985	19.1255	18.7766	18.0418	17.3174	15.6446	10.4166	7.4849	0.9899
Filtered

TABLE 4
SNR comparisons of Lenna image with Gaussian impulse noise N(150, 1)
Lenna150	10%	20%	30%	40%	50%	60%	70%	80%	90%	100%
Source	10.4628	7.4211	5.6981	4.4827	3.5464	2.7434	2.0715	1.4676	0.9646	0.5199
Image
Median	22.4472	20.0665	16.4491	13.3095	10.2374	7.4684	5.2682	3.3371	1.915	0.7521
Filtered
WFM	20.3724	18.8613	17.0277	15.4704	13.7193	11.4278	8.9774	6.0738	2.959	0.6548
Filtered
Phase I	20.2181	18.9669	17.2263	15.6029	13.7321	11.7205	9.1009	6.0836	0.0019	0
Filtered
Phase II	19.9616	19.1947	18.6209	18.0375	16.9158	15.3521	12.3493	8.6377	0.0001	−0.0001
Filtered
Phase III	19.7374	18.956	18.6085	18.174	17.4214	16.1191	13.348	9.7365	0.0001	−0.0001
Filtered

TABLE 5
SNR comparisons of Pepper image with Gaussian impulse noise N(100, 1)
peppers100	10%	20%	30%	40%	50%	60%	70%	80%	90%	100%
Source	12.9438	10.0034	8.2382	6.9488	6.0112	5.2064	4.5462	3.9724	3.4736	3.0067
Image
Median	24.0664	21.5265	18.8431	15.529	12.6623	10.0499	7.9544	6.1696	4.7212	3.5189
Filtered
WFM	21.0013	19.6803	18.498	16.7943	15.2045	13.4744	11.599	9.1467	6.2311	2.9955
Filtered
Phase I	17.4986	19.9714	18.6179	13.2332	15.2856	11.8912	11.2605	9.0722	6.2196	3.1031
Filtered
Phase II	17.988	19.8917	19.1558	14.0177	17.5211	13.4005	13.5415	11.4211	6.9028	3.1005
Filtered
Phase III	18.0673	19.5867	18.9685	14.1328	17.7265	13.6984	14.0165	12.4223	7.1967	3.1059
Filtered

TABLE 6
SNR comparisons of Pepper image with Gaussian impulse noise N(150, 1)
peppers150	10%	20%	30%	40%	50%	60%	70%	80%	90%	100%
Source	10.5023	7.5439	5.7997	4.5544	3.6344	2.7545	2.146	1.5176	1.0333	0.5724
Image
Median	23.3079	20.5331	17.0017	13.3635	10.3423	7.3407	5.4003	3.4215	2.03	0.8134
Filtered
WFM	21.1766	19.7254	17.6346	15.444	13.3504	10.6133	8.7324	4.1155	2.6608	−0.1175
Filtered
Phase I	20.6664	12.9899	17.3114	10.8172	13.3292	10.5972	8.7219	5.8571	2.9736	−0.3347
Filtered
Phase II	20.1187	13.492	18.8707	12.1188	16.9852	14.4236	12.3534	8.1697	3.7543	−0.3553
Filtered
Phase III	19.7929	13.533	18.7827	12.281	17.5308	15.5836	13.7786	9.224	4.1097	−0.3545
Filtered

TABLE 7
SNR comparisons of Boat image with Gaussian impulse noise N(100, 1)
boat100	10%	20%	30%	40%	50%	60%	70%	80%	90%	100%
Source	14.0246	11.0775	9.2924	8.0422	7.1127	6.3088	5.6149	5.0504	4.5253	4.0722
Image
Median	20.9373	19.6684	17.8279	15.5728	13.33	10.9758	8.9781	7.3076	5.8568	4.6913
Filtered
WFM	19.0911	18.2913	17.3357	16.4812	12.5724	12.9488	11.7269	9.8422	7.1108	4.3928
Filtered
Phase I	19.1426	18.2104	17.4986	16.2999	15.0153	13.5299	11.8167	9.7533	7.1591	4.0939
Filtered
Phase II	18.6621	18.0138	17.7345	17.0894	16.3519	15.2701	13.728	11.2164	7.6578	4.046
Filtered
Phase III	18.2269	17.7008	17.4702	16.9446	16.3657	15.5164	14.1773	11.8487	7.8542	4.0377
Filtered

TABLE 8
SNR comparisons of Boat image with Gaussian impulse noise N(150, 1)
boat150	10%	20%	30%	40%	50%	60%	70%	80%	90%	100%
Source	11.585	8.5414	6.8766	5.6137	4.5991	3.8314	3.1981	2.6063	2.0628	1.639
Image
Median	20.65	18.7815	16.5899	13.7566	10.8475	8.3562	6.3518	4.536	3.0001	1.9455
Filtered
WFM	9.4359	8.9687	16.7299	15.1802	13.2572	11.3381	9.2217	6.5049	3.4561	0.6624
Filtered
Phase I	18.9864	16.7837	15.1843	13.7756	12.5597	11.511	10.7019	8.9552	8.2656	6.7594
Filtered
Phase II	18.6295	17.0105	15.4449	14.0047	12.7847	11.7693	10.9927	8.9546	8.3607	6.7828
Filtered
Phase III	18.2229	16.972	15.4704	14.0772	12.8166	11.7831	10.9909	8.9546	8.3006	6.7881
Filtered

TABLE 9
SNR comparisons of Baboon image with Gaussian impulse noise N(100, 1)
Baboon100	10%	20%	30%	40%	50%	60%	70%	80%	90%	100%
Source	18.4758	15.5736	13.7707	12.5765	11.5704	10.7777	10.116	9.5356	9.0272	8.5678
Image
Median	23.6095	22.9403	21.5209	19.7481	17.5821	15.5866	13.6753	12.037	10.6483	9.4785
Filtered
WFM	22.4393	21.6793	20.778	19.8066	18.9528	17.8874	16.5002	13.8104	11.1789	8.6852
Filtered
Phase I	21.9419	21.1672	20.1341	19.9752	19.1635	18.1297	16.6143	14.2655	11.367	9.9405
Filtered
Phase II	22.0161	21.3949	20.6081	21.1954	20.8725	20.2771	19.1193	16.2114	11.8374	10.0271
Filtered
Phase III	21.8863	21.3147	20.5594	21.1924	20.9886	20.594	19.6867	16.9031	12.0233	10.0546
Filtered

TABLE 10
SNR comparisons of Baboon image with Gaussian impulse noise N(150, 1)
Baboon150	10%	20%	30%	40%	50%	60%	70%	80%	90%	100%
Source	16.3252	13.2688	11.4904	10.2338	9.2782	8.4937	7.7843	7.2322	6.73	6.2668
Image
Median	23.5175	22.4221	20.5665	18.0101	15.4198	13.0346	10.8073	9.1397	7.746	6.5356
Filtered
WFM	23.0807	22.4737	21.5114	20.1828	18.5514	16.8628	14.4697	11.972	8.9851	5.858
Filtered
Phase I	22.0338	21.4196	20.3983	19.591	18.2496	16.6579	14.9453	11.9925	9.0737	8.5788
Filtered
Phase II	22.1727	21.7479	21.1863	21.1919	20.5587	19.7956	18.1092	14.4447	9.8787	8.7907
Filtered
Phase III	22.0776	21.7009	21.1578	21.2538	20.8639	20.3329	19.013	15.456	10.2153	8.8242
Filtered

TABLE 11
PSNR comparisons of Lenna image with Gaussian impulse noise N(100, 1)
Lenna100	10%	20%	30%	40%	50%	60%	70%	80%	90%	100%
Source	18.6353	15.7069	13.9699	12.7481	11.7741	10.9825	10.3064	9.7536	9.2309	8.7831
Image
Median	28.9169	26.9576	24.4145	21.2366	18.4348	15.8571	13.7444	11.9891	10.4961	9.3344
Filtered
WFM	26.0068	24.6781	23.3802	21.8806	20.6351	19.0108	17.2749	14.9147	11.9451	8.772
Filtered
Phase I	26.5968	25.6125	24.3839	22.8391	21.4447	19.5914	17.6008	13.9367	12.0631	6.6943
Filtered
Phase II	26.159	25.646	25.2025	24.689	23.7713	22.6573	20.6722	15.6624	12.9323	6.8207
Filtered
Phase III	25.883	25.36	24.9871	24.6381	23.9034	23.1789	21.5062	16.2781	13.3465	6.8514
Filtered

TABLE 12
PSNR comparisons of Lenna image with Gaussian impulse noise N(150, 1)
Lenna150	10%	20%	30%	40%	50%	60%	70%	80%	90%	100%
Source	16.3243	13.2826	11.5596	10.3442	9.408	8.6049	7.9331	7.3291	6.8262	6.3814
Image
Median	28.3087	25.9281	22.3106	19.1711	16.099	13.33	11.1298	9.1986	7.7765	6.6136
Filtered
WFM	26.2339	24.7228	22.8892	21.3319	19.5809	17.2894	14.8389	11.9354	8.8205	6.5164
Filtered
Phase I	26.0796	24.8285	23.0879	21.4645	19.5936	17.5821	14.9624	11.9451	5.8635	5.8615
Filtered
Phase II	25.8232	25.0562	24.4824	23.899	22.7773	21.2136	18.2108	14.4993	5.8616	5.8615
Filtered
Phase III	25.599	24.8175	24.47	24.0355	23.283	21.9806	19.2095	15.5981	5.8616	5.8615
Filtered

TABLE 13
PSNR comparisons of Pepper image with Gaussian impulse noise N(100, 1)
peppers100	10%	20%	30%	40%	50%	60%	70%	80%	90%	100%
Source	18.8467	15.9063	14.1411	12.8516	11.9141	11.1092	10.449	9.8753	9.3765	8.9096
Image
Median	29.9692	27.4294	24.7459	21.4318	18.5652	15.9527	13.8572	12.0724	10.6241	9.4218
Filtered
WFM	26.9042	25.5832	24.4008	22.6971	21.1074	19.3773	17.5019	15.0496	12.134	8.8984
Filtered
Phase I	23.4015	25.8742	24.5208	19.1361	21.1884	17.7941	17.1634	14.975	12.1224	9.006
Filtered
Phase II	23.8909	25.7945	25.0586	19.9206	23.424	19.3033	19.4443	17.3239	12.8057	9.0033
Filtered
Phase III	23.9702	25.4895	24.8713	20.0356	23.6294	19.6012	19.9193	18.3251	13.0996	9.0087
Filtered

TABLE 14
PSNR comparisons of Pepper image with Gaussian impulse noise N(150, 1)
peppers150	10%	20%	30%	40%	50%	60%	70%	80%	90%	100%
Source	16.4052	13.4467	11.7026	10.4573	9.5373	8.6574	8.0488	7.4204	6.9361	6.4753
Image
Median	29.2107	26.436	22.9045	19.2664	16.2452	13.2435	11.3031	9.3243	7.9328	6.7163
Filtered
WFM	27.0794	25.6282	23.5374	21.3469	19.2533	16.5162	14.6353	10.0183	8.5636	5.7853
Filtered
Phase I	26.5692	18.8928	23.2143	16.72	19.2321	16.5	14.6248	11.76	8.8765	5.5682
Filtered
Phase II	26.0215	19.3948	24.7735	18.0217	22.8881	20.3265	18.2562	14.0726	9.6571	5.5476
Filtered
Phase III	25.6957	19.4358	24.6855	18.1838	23.4336	21.4864	19.6814	15.1268	10.0125	5.5484
Filtered

TABLE 15
PSNR comparisons of Boat image with Gaussian impulse noise N(100, 1)
boat100	10%	20%	30%	40%	50%	60%	70%	80%	90%	100%
Source	18.8728	15.9257	14.1406	12.8904	11.9609	11.157	10.4631	9.8986	9.3735	8.9204
Image
Median	25.7855	24.5166	22.6761	20.421	18.1782	15.824	13.8263	12.1558	10.705	9.5395
Filtered
WFM	23.9393	23.1395	22.1839	21.3294	17.4206	17.797	16.5751	14.6904	11.959	9.241
Filtered
Phase I	23.9908	23.0586	22.3468	21.1481	19.8635	18.3781	16.6649	14.6015	12.0073	8.9421
Filtered
Phase II	23.5103	22.862	22.5827	21.9376	21.2001	20.1183	18.5762	16.0646	12.506	8.8942
Filtered
Phase III	23.0751	22.549	22.3184	21.7928	21.2139	20.3646	19.0255	16.6969	12.7024	8.8859
Filtered

TABLE 16
PSNR comparisons of Boat image with Gaussian impulse noise N(150, 1)
boat150	10%	20%	30%	40%	50%	60%	70%	80%	90%	100%
Source	16.4332	13.3897	11.7248	10.4619	9.4473	8.6796	8.0463	7.4545	6.911	6.4872
Image
Median	25.4982	23.6297	21.4382	18.6048	15.6958	13.2044	11.2	9.3842	7.8483	6.7937
Filtered
WFM	14.2841	13.8169	21.5781	20.0284	18.1054	16.1864	14.0699	11.3531	8.3044	5.5106
Filtered
Phase I	23.8346	21.6319	20.0325	18.6238	17.4079	16.3592	15.5501	13.8035	13.1138	11.6076
Filtered
Phase II	23.4777	21.8587	20.2931	18.8529	17.6329	16.6175	15.8409	13.8028	13.2089	11.631
Filtered
Phase III	23.0712	21.8202	20.3187	18.9254	17.6648	16.6313	15.8391	13.8028	13.1488	11.6363
Filtered

The above tables contain data of two phases' filtering results. A further processing called the Phase III can be performed by applying the same Phase II operations over the image again. We see that AMFG is stable since repeating Phase II operations will not reduce the performance. From the tables, the performance of Phase I is close to those of WFM but after Phase II, the performance is 3 to 4 dB better than those of WFM. Even 5 dB better performances can be found when the impulse mean strength is over 200.

In summary, this invention adopts two-phase image filter. In the first phase, most noises are removed. For noise ratio higher than 30%, there would be corrupted spots (a small region of corrupted pixels) left. To make the filtering result be much more close to the original intact image, we perform thresholding over the filtered image of the first phase and utilize the thresholded image for corruption determination of pixels and for further filtering of noises as well. According to the experiments, AMFG filter evidently increase the SNR and PSNR performances. Moreover, in addition to filtering, at the same time we can perform image analyses such as thresholding, edge detection, and image segmentation. The total time and spatial complexity of computation is as following table and we can find that they are much lower than those of the state-of-art analysis algorithms.

TABLE 17
Spatial complexities of AMFG filter and proposed analyses operations.
Phase I Filter     N = image border length; L = gray levels
Maximum memory     N2 + L
usage
Total memory usage 3N2
Phase II Filter    N = image border length; n = sample window
                   border length
Maximum memory     N2
usage
Total memory usage 3N2

TABLE 18
Time complexities of AMFG filter and proposed analyses operations.
Phase I Filter	N = image border length; n = sample window border length; cn = the number of sections;
	cp = the average number of gray levels in a section; L = the total number of gray levels in the
	histogram and L = cn × cp; k2 = the number of fuzzy partitions (Thresholds).
Assignments	6(N2 × k2 × n2) + 6(N2 × k2) + 8(cn + k2) + 2(N2 × k1) + 11N2 + 3cn + 764	O(N2 × k2 × n2 + L)
Additions	5(N2 × k2 × n2) + 2(N2 × k1) + 2(N2 × k2) + 2(4cn + 3k2) + N2 + 3cn − 4	O(N2 × k2 × n2 + L)
Multiplications	(N2 × k2 × n2) + (cn × cp) + N2 + cn + 2	O(N2 × k2 × n2 + L)
Divisions	2(N2 × k2 × n2) + (N2 × k2) + N2 + 2cn + 1	O(N2 × k2 × n2 + L)
Comparisons	2(N2 × k2 × n2) + 2(N2 × k1) + 2N2 + 3(cn + cp − 1) + k2	O(N2 × k2 × n2 + L)
Floor	cn + 1	O(cn)
operations
Total	O(N2 × k2 × n2 + L)
Phase II Filter	N = image border length; n = sample window border length;
Assignments	[5(n2 − 1) + 3](N − 2)2	O(n2 × N2)
Comparisons	(n2 − 1)(N − 2)2 + 2n(N − 2)2 (Median operations: 2(N − 2)2)	O(n2 × N2)
Increments	(n2 − 1)(N − 2)2	O(n2 × N2)
Floor	[2(n2 − 1) + 2](N − 2)2	O(n2 × N2)
operations
Total	O(n2 × N2)

In the above complexity evaluation, finding thresholds in the histogram requires 10c_n+8k₂−4 memory accesses, L+15c_n+6k₂+3c_p−7 additions, one division, 2 multiplications, and one floor operation. That is, the complexity of finding thresholds is of order O(L) which is much lower than state-of-art methods. With such a low cost as Tables 17 and 18, this invention effectively and efficiently performs image analyses such as thresholding, edge detection, and segmentation. This invention breaks through the dilemma that heavily corrupted images cannot be analyzed. Therefore, this invention can be widely applied in industries such as medical imaging, image compression, multimedia communications, space technology, . . . , etc. This invention possesses wide applicability and high industrial worth. Especially, the thresholded image TH in Phase I can also be replaced with result images from other thresholding algorithms and becomes the input of Phase II. This compatibility enlarges the application area. Therefore, this invention does achieve the objective and there is no other similar methods found in open area. This invention meets the required criteria to apply for the patent.

Claims

1. An analysis method comprising a Phase I and a Phase II image processing, wherein the phase I processing producing a histogram of a noisy image;according to the histogram, automatically generating a fuzzy partitions represented by LR fuzzy sets and parameters of the fuzzy sets' membership functions; wherein the parameters are left spread α, mean m, and right spread β such that the fuzzy partitions are represented by LR fuzzy sets (α, m, β)LR;with these automatically generated fuzzy sets and their membership functions, using fuzzy inference as noise filter and producing the Phase I filtered and thresholded images;the Phase II processing scaning the resultant images of Phase I by sample windows and comparing the centered pixel with surrounding pixels in a sample window of the thresholded image to determine whether the center pixel is a part of a corrupted spot;in a sample window, the non-corrupted pixels being used for filling the center pixel that is corrupted;the Phase II analysis again applying the same thresholding algorithm in Phase I to obtained the Phase II thresholded image; wherein the Phase II thresholded image is used to perform the edge detection analysis.

2. The method as claimed in claim 1, wherein analyses of heavily corrupted images using fuzzy partitions as claimed in claim 1 comprises the steps of: regarding the image as a N-by-N matrix, and calculating a histogram thereof;cutting the histogram into several sections and summing up the statistics in every section;if the summation of previous section is negative and present one is positive, determining that the starting gray level of present section being a coarse valley point;between any two consecutive coarse valley points, finding out the gray level with the maximum statistic as a peak point and between any two peak points, finding out the gray level with the minimum statistic as an accurate valley point;regarding a starting point and an end point of the histogram as two valley points; one peak point and two neighboring valley points forming a fuzzy partition; wherein the peak point is the mean m of the corresponding LR fuzzy set (α, m, β)LR; a distance from the left neighboring valley point to the peak point is the left spread parameter α while the distance from the right neighboring valley point to the peak point is the right spread parameter β;for each of the fuzzy partition (α, m, β)LR, calculating membership grades of all the pixels in the sample window; using the membership grade as weighting, calculating the weighted average of each sample window; andcalculating a maximum likelihood estimator of each sample window; choosing the weighted average that is the most close to the estimator as the output so as to obtain the filtered image of Phase I.

3. The method of claim 2, wherein the histogram's starting point is taken at gray level 5 and the end point is taken at the maximum gray level L minus 5, that is L−5. In this way most impulse noises will not attend the calculations of fuzzy partitions, filtering, and thresholding.

4. The method of claim 2, wherein the Phase I thresholded image is obtained by replacing all pixels' gray levels with their corresponding fuzzy partition's mean m.

5. The method of claim 2, wherein the membership grade of each pixel multiplying with each pixel's gray level is the weighted value; in each sample window, summing up all the weighted value to obtain the weighted sum; in a sample window, the weighted sum over the sum of membership grades is the weighted average.

6. The method of claim 2, wherein the weighted average that is the most close to the maximum likelihood estimator of the pixels in every sample window is selected to fill into a matrix until the scanning is end. When the scanning is end, the Phase I filtered image is obtained.

7. The method of claim 1, wherein the steps of Phase II are: putting the thresholded image of Phase I in a N×N matrix TH(I, J);horizontally scanning the pixels of TH(I, J) by a n×n sample window xd(I+i,J+j);setting threshold value Kt for corruption determination and comparing the center pixel of the sample window with its surrounding pixels;if there are fewer than Kt pixels having the same gray levels as the center pixel, the center pixel being regarded corrupted; otherwise, move the sample window to the next position and go back to (2);calculating the median of the non-corrupted pixels in the sample window, which is different from the center pixel;replacing the center pixel's gray level with the median;obtaining the Phase II filtered image after that each sample window is processed with the above procedure;in this way, after two-phase filtering, the AMFG filter evidently increasing the SNR and PSNR values such that the image are clear.

8. The method of claim 7, wherein in the fourth step of the Phase II supposes that the threshold for corruption determination is Kt and the accumulated number of corrupted pixels is K; comparing the center pixel with its surrounding ones pixel by pixel in the sample window, increasing K if a neighbor pixel has the same gray level as the center one; otherwise, remaining the value of K; when K<Kt, the center pixel is regarded as part of a corrupted spot otherwise when K≧Kt, it is regarded as intact.

9. The method of claim 1, wherein the Phase II thresholded image is obtained by applying the thresholding method in claim 4 over the Phase II filtered image.

10. The method of claim 1, wherein the Phase II thresholded image is scanned using sample window and if the center pixel is different from one of the left, upper left, upper, and upper right pixel, the maximum gray level is put in the corresponding position of a matrix ED, otherwise 0 is put in that position; after every the pixels of the Phase II thresholded image is scanned and processed in this way, the edge-detected image ED is obtained where all the edges are white.

11. The method of claim 1, wherein the Phase II thresholded image is scanned using sample window and if the center pixel is different from one of the left, upper left, upper, and upper right pixel, 0 is put in the corresponding position of a matrix ED, otherwise the maximum gray level is put in that position; after every the pixels of the Phase II thresholded image is scanned and processed in this way, the edge-detected image ED is obtained where all the edges are black.

12. The method of claim 1, wherein the horizontal scanning of an image is from upper left to lower right, from upper right to lower left, from lower left to upper right, and from lower right to upper left.

13. The method of claim 1, wherein the horizontal scanning of an image is replaced with vertical scanning.

14. The method of claim 1, wherein the size of the sample windows is 3×3 and additionally the size is changed according to the situation of noise corruption; and the sample window size is 5×5, 7×7, 5×7, and 7×5 for filtering larger corrupted spot.

15. The method of claim 1, wherein a reference value v for cutting histogram into sections is determined according to Nyquist Rate sampling theorem so that the number of sections cn is calculated.

16. The method of claim 1, wherein the Phase I thresholded image TH is replaced with any other thresholded image by other thresholding methods and becomes input of Phase II for further filtering and analyses.

17. The method of claim 1, wherein the Phase II is applied more than once and thus called Phase III, Phase IV, . . . , etc; and the PSNR and SNR will not change obviously in these later phases.