TWI273514B - Analyses of heavily corrupted images using fuzzy partitions - Google Patents

Abstract

This invention based on 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. This invention break through the bottleneck that heavily corrupted image cannot be analyzed.

Description

1273514 IX. Description of the invention: [Technical field of the invention] The present invention utilizes the principle of a fuzzy brake to enable an image filled with Gaussian and pulsed mixed noise to self-organize to obtain its original image feature values, which will be It is represented by a fuzzy set, and the attribute function of the fuzzy set can be used to perform a sigma hole; multi-level thresholding and segmentation, image edge capture (edge detecti〇) n) and other functions, and after the noise filtering and multi-valued image, can be used as the second stage of the noise filtering reference, after the two stages of filtering noise processing can greatly increase the image The SNR and pSNR values allow the image noise filter to make the image sharper. And re-obtain the image after processing such as multi-value, segmentation, and edge capture. The image after segmentation can be applied for medical image diagnosis 'MPEG compression, image recognition and other fields. [Previous technology] Shirts are the most direct and effective way to convey messages. Especially in the technology age, the application of image transmission has been widespread in our lives, such as medical, security, space and other technologies, but when images are intercepted or In the process of transmission through the medium, it may be because the resolution of the interception device is too low, the compression technique of the storage format is too bad, and the pulse and Gaussian noise generated by the defective channel or the noise-containing pipeline make the original clear. The image produces color that is tender and blurred, destroying the image and degrading the industrial value. Therefore, before the image is presented to the producer, image processing reduction techniques are usually required to further perform image analysis and calculations such as image multi-value (image) 1273514 multi-level thresholding) and image segmentation (image segmentatl〇n). In the field of image processing, there are many existing filtering methods for noise removal, such as: low pass filters > median f i Iter > weighted median f i Iters (WM,

Browrigg, 1984) > center weighted median filters(CWM, Ko and

Lee, 1991)··· and other methods. Most of the methods proposed in the past are only a noise filtering action on the contaminated image. After a noise filtering operation, the image still has a large percentage of the image that cannot be removed. The noise is not perfect. For example, Median Filter will perform intermediate value calculation on its window regardless of whether there is any noise in the sampling window. When the image noise exceeds 30%, the noise filtering effect will be rapidly deteriorated, and it will be easy to The part of the noise is also included in the basis of the interim value calculation, which will make the already contaminated image more blurred. For example, when WFMFilter processes images, the gray scales of the images are fixed into dark, medium and bright. It is impossible to automatically obtain the grayscale distribution points according to the characteristics of the images, so as to produce better results, and the visual effects are not required by Median Filter. It's good, because when the proportion of noise is low, there may be only one or two noise points in the sampling window of a nine-square grid. If you use WFMFilter to process, there is a tendency to blur the lines on the processed image. In the field of image analysis, there are three commonly used processes, Thresholding segmentation and edge detection, for images with high pollution (contamination ratio greater than 3〇%). Fast (low computational complexity) and efficient methods for criticalization, segmentation, and edge detection have not been found in the prior art. Therefore, in terms of progress, the present invention is actually a large 1273514 • breakthrough. SUMMARY OF THE INVENTION The main object of the present invention is to provide a two-stage image filter capable of removing high proportion lions, and to perform criticalization, segmentation (segmentat problem), and edge_(_detection). ) and so on. In addition to obtaining more realistic raw images and performing effective image analysis, image processing time can be significantly reduced. The main feature of the a tree shape is to provide a two-stage filter and wave filter. The image of the first segment is filtered to generate a histogram, and a histogram of the noise image is used. The LR parameter Left spread α, Mean m, and Right Spreacj of the LR parameter that automatically generates the membership function (Membership 1111) are used to represent their LR 集合 集合 collection (α , m, point) LR, then use the automatically obtained fuzzy set and the membership function to filter the noise according to the fuzzy theory and generate the first stage of the Thresholded Image. In the second stage of the image filter, the image obtained in the first stage is resampled and converted into a less complex image. At this stage, the action of the £& boundary image generated by the previous Deng knife will be sampled. The center point of the sampling is compared with the values around it to determine whether the first stage has not filtered out the noise points, and the part of the non-noise point in the sampling view is used as the basis for filling the center point. [Embodiment] The present invention is directed to a preferred embodiment of the present invention in order to achieve the above-mentioned purpose and effect, and is illustrated in the accompanying drawings. Field _, the way to filter out image noise and the general communication system 7 1273514 is the concept of chopping, both of which are to sample, transfer, filter or give the input signal in sequence. The common processing of the square cut is to first process each pixel through a ship shovel sampling 彳 « (the coffee pot wind tank on the image like a CRT display ϋ horizontal scanning way to process each pixel, each time interval - pixels (4)). For example, the image-sampling window is horizontally scanned and moved. Move the sample window from left to right, and when the silk moves _ the rightmost (four), the bribe returns to the leftmost side and will look down the crane - a pixel, and Wei performs all the pixel processing. ^ The schematic diagram of the present invention also uses fuzzy inference to perform image processing. As shown in the second figure, the size of the sampling window is determined by design specifications and algorithms. The sampling window of the frame is most often used as a sampling window of a 9-square grid (3 X 3). Therefore, in the sampling window, the sampling window with odd values is better. In the odd sampling window, the sampling of the nine squares is the best. Because the image is detected by continuous signals, it is sampled in the image of the nine squares. When there is no noise involved, the difference in pixel values around the middle point 不会 is not too large. The pixels in the sampling window are subjected to a fuzzy inference system to produce a new pixel gray instead of the original pixel of the sampled towel. The present invention enables a high-soil-like image analysis method to include a two-phase filter. The description is as follows: 1. Phase I method In the image wave of the first phase (Phase I), the fuzzy segmentation method is used to automatically obtain the membership function of the fuzzy segmentation (Μ-φ FUnCti〇n) parameter α according to the image characteristics. After m and β, the noise is filtered according to the fuzzy theory. The filtering effect of 1273514 is much better than the intermediate value and the processing effect using the fixed membership function. In addition, due to the use of fuzzy segmentation, if the image filtered by the first stage is classified by a single grayscale value in each segment, that is, when the grayscale value belongs to the same segmenter, it is represented by the same grayscale value. Achieve the purpose of multi-valued in the first phase. Please refer to the steps in the third figure as follows: (1) Statistics generate a histogram (hist〇gram) and segment the histogram. Substituting the contaminated image of NxN pixels into the matrix of XA (i, the histogram hs(x) of the image matrix XA(i,j) is counted, and the histogram value of 5~25〇 is retained, so that You can first filter out most of the ImpUlse N〇ise in the image. Re-count the correct starting and ending points of the histogram, and compare the value of each grayscale of the histogram with 0. When the _th and last _ are not the gray scale value of Q, the effective starting value hss and the end point value hsf in the matrix of hs(x) are calculated for the correct starting and ending points of the histogram, and the histogram The value of each gray scale is compared with 〇. When the value is compared to the first and last gray scale values that are not 〇, it is the correct start and end of the gamma shed for finding the hs(x) The membership function (Mefflbership Function), so we must downsample the hs(x) as (1)(2), where y is the sampling reference value 'for the number of sampling segments, and the Q riding segment samples the gray level of the towel. c. =( 0 X 2 + 1) + 2)x2 0) c a « Ρ . (2) Find the local valley point from the mother # again sampling (the lowest gray level of the local minimum ten words) 12 73514 and peak point (gray scale of the local maximum statistic). Calculate the position of each segment of hs(x) 7· (0) and the average value rhS(i) of each sample segment, as shown in (3)(4). c,,〇= ΜχΜχ/

Lv ^ J _ (3) (4) Σ hsix) ^(0 = ^^——

CP

Calculate the average difference value r for each sampled segment. (1), as shown in (5). "(5) Judging whether 1"(1) and r<5(i-1) have a preset valley Cm(i), finding a peak between two preset valleys Q(1) and must-1) Point CM(1), as shown in (6)(7). {cw (k)\keN}^ {chs {i)\r〇{i)r〇(/ -1) < 〇5 / € ^ {1}j cm(0 = Max hs(x) () ^ (7) _ Find the correct valley point Cm(i) between the peak points CM(1) and CM(H), as in (8). (8) (3) Calculate each segment of fuzzy segmentation· The parameters of the 夂" (4), points, as shown in (9)(10)(11), to calculate the membership function for each segmentation. Left spread parameters: (9) α/ =^(0 -^(0 mean parameter: mi =^(0 (10) 10 1273514 'right spread parameter · + (11) We use (ίη, 仙, to represent the fuzzy segmentation. Use sh X & sampling windowΥΑ To sample the ΧΑ, for the (h, w)-th sampling window as shown in (12). ie[h,hJr(sh -1)] ((7) je[wyw^(sw -1)]

Substituting YA(i,j) into the third segment fuzzy segmentation (Jk, fflk, i^k) LR to obtain the kth group membership degree of the pixel YA(i,j)...(i,j), such as (13) . 'nheUjUj) 1 YA(Uj)Qmk, ak Pk , where aQb = mdx(a-b, 0) (13)

(d) Calculate the maximum possible estimate and weighted average of the sampling window. Calculate the maximum possible estimated value AVd〇(i4) in the Sh X Sw sampling window YA, multiply YA(i,j); ca(i,j) obtain the weighted value Wk, ux as (15), and divide by The wk of the membership function, u is (16), that is, the weighted average is taken as (10) as shown in (17). ΣΣ^;) AV^±Uil-- /=1 y=l Wkyu = /=1 j=\ n,(10)Ten (Ά,")+0.5)" (14) (15) (16) 11 (17) 1273514 • (5) Output the filtered result and threshold the image. Calculate the difference between the k-th weighted average Wk, ain and the maximum possible estimate AVk. The closest weighted average Wtgm is shown in (18), and it is selected to output to 〇K(h, w). This process is completed until the NxN image XA is completely sampled, that is, h := N iw = N. The matrix οκ is the first stage filtered image. v丨;卜户μ卜'4 — 08)

After 0K(h, w) is criticalized, it becomes a criticalized image matrix training, as shown in (19). (V/ e iV - {1),^(/) < 0K(h^w) < cm(in))[TH(Kw) = c,v/ (/)] (19) TH is the first A staged image (thresholded image). 2. Phase II method In the image filter of the second stage, the image of the original gray-scale complexity is converted into a lower-complexity image by the Thresholding action, so when in the sampling window, the phase If the neighboring pixel values are different, it means that the edge of the image may be here, but if the number of different points exceeds a certain number, the possibility that the first stage does not filter out the noise is greatly improved. Therefore, at this stage, the action of sampling the critical image generated in the first stage is compared, and the center point of the sample is compared with the value of the periphery. If the center point is the same as the surrounding value, the noise point is judged. The value K is the cumulative action. When there is one point, the total is incremented by 1. When the sampling center point is compared with the surrounding points, and the noise point judgment value K is also smaller than the set noise threshold value Kt, the sampling is performed. The center point of the window is determined as the first order 12 1273514 segment of the filter, and the non-noise portion of the sampling window is used as the basis for filling the center point. The method steps are as follows (please refer to the fourth figure): (1) The first stage critical image is placed in an Nxiv TH matrix. (2) Horizontal scanning is performed on the TH matrix using the sampling window xd of /7X77, and the horizontal scanning is performed on the sampling window X0 of the 77X knife for the 〇κ matrix, and the (I, J)-th sampling window is (20).

X〇0,J) = 〇K(U) [Λ/+(,7-1)] (3) Set the noise threshold Kt. Compare the center point of the sampling window with the surrounding points. The values are the same, and the hysteresis judgment value κ is accumulated. The values are not the same, and the noise judgment value K does not change. (4) Comparing the judgment value K of the noise point and the threshold value Kt of the noise determination. K>Kt, the center point is not considered to be noisy. And move this (I, j)-th sampling window to the next pixel.

/ Outside K<Kt, the center point is determined to be a noise point. And carry out the following steps (5) (5) Take out the sampling window xo deducting the center point χ ο (/ + 卜 non-noise pixel grayscale value; and do the intermediate value operation. (/,) will be judged as a xo (/+""/2"j,J+b,2} is replaced by the value of the middle and the back. When all the sampling is completed, that is, I=N and: ί=Ν, QK scale update ~ Yunjin is Yuancheng One-stage noise removal procedure. ^ 13 1273514

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

Lenna impulse noise 100 SNR comparison table LennalOO, 10% 20% 30% 40%; 50% 60% 70% 80% 90% 100% Source Image 12.7738 9.8454 8.1084 6.8865 5.9126 5.1209 4.4449 3.8921 3.3693 2.9216 Median Filtered 23.0554 21.0961 18.553 15.3751 12.5733 。 。 。 。 。 。 。 。 。 。 。 19.1255 18.7766 18.0418 17.3174 15.6446 10.4166 7.4849 0.9899

Lenna impulse noise 150 SNR comparison table Lennal50 10% 20% :,30% 113⁄4 '50% / 60% :70% 80% 90% 100% Source Image 10.4628 7.4211 5.6981 4.4827 3.5464 2.7434 2.0715 1.4676 0.9646 0.5199 Mediian: Filtered, 22.4472 20.0665 16.4491 13.3095 10.2374 7.4684 5.2682 3.3371 1.915 0.7521 WFM Filtered , 20.3724 18.8613 17.0277 15.4704 13.7193 11.4278 8.9774 6.0738 2.959 0.6548 Phase I Filtered 20.2181 18.9669 17,2263 15.6029 13.7321 11.7205 9.1009 6.0836 0.0019 0 Phase II Filtered 19.9616 19.1947 18.6209 18.0375 16.9158 15.3521 12.3493 8.6377 0.0001 - 0.0001 Phase III Filtered 19.7374 18.956 18.6085 18.174 17.4214 16.1191 13.348 9.7365 0.0001 •0.0001

Peppers impulse noise 100 SNR comparison table pepperers 100 10% .20% 30% __ 60% 70% 80% 90% 1003⁄4 Source Image 12.9438 10.0034 8.2382 6.9488 6.0112 5.2064 4.5462 3.9724 3.4736 3.0067 Median Filtered 24.0664 21.5265 18.8431 15.529 12.6623 10.0499 7.9544 6.1696 4.7212 3.5189 WFM Filtered 21.0013 19.6803 18.498 16.7943 15.2045 13.4744 11.599 9.1467 6.2311 2.9955 Phase I Filtered 17.4986 19.9714 18.6179 13.2332 15.2856 11.8912 11.2605 9.0722 6.2196 3.1031 Phase II Filtered 17.988 19.8917 19.1558 14.0177 17.5211 13.4005 13.5415 11.4211 6.9028 3.1005 Phase III Filtered 18.0673 19.5867 18.9685 14.1328 17.7265 13.6984 14.0165 12.4223 7.1967 3.1059 15 1273514

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

SNR of Boat impulse noise 100 Comparison table boatlOO 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Source Image 14.0246 11.0775 9.2924 8.0422 7.1127 6.3088 5.6149 5.0504 4.5253 4.0722 Median Filtered 20.9373 19.6684 17.8279 15.5728 13.33 10.9758 8.9781 7.3076 5.8568 4.6913 WFM Filtered 19.0911 18.2913 17.3357 16.4812 12.5724 12.9488 11.7269 9.8422 7.1108 4.3928 Phase I Filtered 19.1426 18.2104 17.4986 16.2999 15.0153 13.5299 11.8167 9.7533 7.1591 4,0939 Phase II Filtered 18.6621 18.0138 17.7345 17.0894 16.3519 15.2701 13.728 11.2164 7.6578 4.046 Phase III Filtered 18.2269 17.7008 17, 4702 16.9446 16.3657 15.5164 14.1773 Ί 1.8487 7.8542 4.0377

SNR of Boat impulse noise 150 Comparison table boatlSO : :10% .20% ' 30% 40% '50% 60% 70% 80% 90% 100% Source Image 11.585 8.5414 6.8766 5.6137 4.5991 3.8314 3.1981 2.6063 2.0628 1.639 MedianFiltered 20.65 18.7815 16.5899 13.7566 10.8475 8.3562 6.3518 4.536 3.0001 1.9455 WFM Filtered 9.4359 8.9687 16.7299 15.1802 13.2572 11.3381 9.2217 6.5049 3,4561 0.6624 Phase I Filtered 18.9864 16.7837 15.1843 13.7756 12.5597 11.511 10.7019 8.9552 8.2656 6.7594 Phase II Filtered 18.6295 17.0105 15.4449 14.0047 12.7S47 11.7693 10.9927 8.9546 8.3607 6.7828 Phase III Filtered 18.2229 16.972 15.4704 14.0772 12.8166 11.7831 10.9909 8.9546 8.3006 6.7881

PSNR Comparison Table for Baboon impulse noise 100 BaboonlOO 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Source Image 18.4758 15.5736 13.7707 12.5765 11.5704 10.7777 10.116 9.5356 9.0272 8.5678 Median Filtered 23.6095 22.9403 21.5209 19.7481 17.5821 15.5866 13.6753 12.037 10.6483 9.4785 WFM Filtered 22.4393 21.6793 20.778 19.8066 18.9528 17.8874 16.5002 13.8104 11.1789 8.6852 Phase I Filtered 21.9419 21.1672 20.1341 19.9752 19.1635 18.1297 16.6143 14.2655 11.367 9.9405 Phase II Filtered 22.0161 21.3949 20.6081 21.1954 20.8725 20.2771 19.1193 16.2114 11.8374 10.0271 Phase III Filtered 21.8863 21.3147 20.5594 21.1924 20.9886 20.594 19.6867 16.9031 12.0233 10.0546 16 1273514

PSNR of Baboon impulse noise 150 Comparison Table Baboonl50 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Source Image 16.3252 13.2688 11.4904 10.2338 9.2782 8.4937 7.7843 7.2322 6.73 6.2668 Median Filtered 23.5175 22.4221 20.5665 18.0101 15.4198 13.0346 10.8073 9.1397 7.746 6.5356 WFM Filtered 23.0807 22.4737 21.5114 20.1828 18.5514 16.8628 14.4697 11.972 8.9851 5.858 Phase I Filtered 22.0338 21.4196 20.3983 19.591 18.2496 16,6579 14.9453 11.9925 9.0737 8.5788 Phase II Filtered 22.1727 21.7479 21.1863 21.1919 20.5587 19.7956 18.1092 14.4447 9.8787 8.7907 Phase III Filtered 22.0776 21.7009 21.1578 21.2538 20.8639 20.3329 19.013 15.456 10.2153 8.8242

Penna comparison table for Lenna impulse noise 100 LennalOO : 10% 20% 30% 40% ; 50% 60% 70% 80% 90% 100% Source Image 18.6353 15.7069 13.9699 12.7481 11.7741 10.9825 10.3064 9.7536 9.2309 8,7831 Median Filtered 28.9169 26.9576 24.4145 21.2366 18.4348 15.8571 13.7444 11.9891 10.4961 9.3344 WFM Filtered 26.0068 24.6781 23.3802 21.8806 20.6351 19.0108 17.2749 14.9147 11.9451 8.772 PKase I Filtered ; 26.5968 25.6125 24.3839 22.8391 21.4447 19.5914 17.6008 13.9367 12.0631 6.6943 Phase II Filtered 26.159 25.646 25.2025 24.689 23.7713 22.6573 20.6722 15.6624 12.9323 6.8207 Phase ffl Filtered 25.883 25.36 24.9871 24.6381 23.9034 23.1789 21.5062 16.2781 13.3465 6.8514

PSNR comparison table for Lenna impulse noise 150_U|I5G 10% 7203⁄4 : 40% 50% <60% ;:;, 70% 80% 90% 100% Source Image 16.3243 13.2826 11.5596 10.3442 9.408 8.6049 7.9331 7.3291 6.8262 6.3814 Median Filtered 28.3087 25.9281 22.3106 19.1711 16.099 13.33 11.1298 9.1986 7.7765 6.6136 WBMFiltered; 26.2339 24.7228 22.8892 21.3319 19.5809 17.2894 14.8389 11.9354 8.8205 6.5164 P6ase ^Filtered 26.0796 24.8285 23.0879 21.4645 19.5936 17.5821 14.9624 11.9451 5.8635 5.8615 iPhase II Filtered 25.8232 25.0562 24.4824 23.899 22.7773 21.2136 18.2108 14.4993 5.8616 5.8615 Phase ΠΙ Filtered 25.599 24.8175 24.47 24.0355 23,283 21.9806 19.2095 15.5981 5.8616 5.8615

Peppers impulse noise 100 PSNR Comparison table pepperslOO 10% 20% 30% m〇>50% 60% 70% 8〇i | 90% Imm Source Image 18.8467 15.9063 14.1411 12.8516 11.9141 i 1.1092 10.449 9.8753 9.3765 8.9096 Median Filtered 29.9692 27.4294 24.7459 21.4318 18.5652 15.9527 13.8572 12.0724 10.6241 9.4218 WFM Filtered 26.9042 25.5832 24.4008 22.6971 21.1074 19.3773 17.5019 15.0496 12.134 8.8984 Phase I Filtered 23.4015 25.8742 24.5208 19.1361 21.1884 17.7941 17.1634 14.975 12.1224 9.006 Phase II Filtered 23.8909 25.7945 25.0586 19.9206 23.424 19.3033 19.4443 17.3239 12.8057 9.0033 Phase III Filtered 23.9702 25.4895 24.8713 20.0356 23.6294 19.6012 19.9193 18.3251 13.0996 9.0087 17 1273514

PSNR of Peppers impulse noise 150 Comparison table peppers 150; 10% 20% 30% 40% i 50% Let 60% 70% 80% 90% 100% Source Image 16.4052 13.4467 11.7026 10.4573 9.5373 8.6574 8.0488 7.4204 6.9361 6.4753 Median Filtered 29.2107 26.436 22.9045 19.2664 16.2452 13.2435 11.3031 9.3243 7.9328 6.7163 WFM Filtered 27.0794 25.6282 23.5374 21.3469 19.2533 16.5162 14.6353 10.0183 8.5636 5.7853 Phase I Filtered 26.5692 18.8928 23.2143 16.72 19.2321 16.5 14.6248 11.76 8.8765 5.5682 Phase II Filtered 26.0215 19.3948 24.7735 18.0217 22.8881 20.3265 18.2562 14.0726 9.6571 5.5476 Phase III Filtered 25.6957 19.4358 24.6855 18.1838 23.4336 21.4864 19.6814 15.1268 10Ό125 5.5484

Boat impulse noise 100 PSNR Comparison table boat 100 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Source Image 18.8728 15.9257 14.1406 12.8904 11.9609 11.157 10.4631 9.8986 9.3735 8.9204 Median Filtered 25.7855 24.5166 22.6761 20.421 18.1782 15.824 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 20.3646 19.0255 16.6969 12.7024 8.8859

PSNR of Boat impulse noise 150 Comparison table boatiSO . /10% 20% _ Figure 40% 50% : 60% 70% 80% 90% 100% Source Image 16.4332 13.3897 11.7248 10,4619 9.4473 8.6796 8.0463 7.4545 6.911 6.4872 Median Filtered 25.4982 23.6297 21.4382 18.6048 15.6958 13.2044 11.2 9.3842 7.8483 6.7937 WFM Filtered 14.2841 13.8169 21.5781 20.0284 18.1054 16.1864 14.0699 11.3531 8.3044 5.5106 Phase IFiltered 23.8346 21.6319 20.0325 18.6238 17.4079 16.3592 15.5501 13.8035 13.1138 11.6076 Phase II Filtered 23.4777 21.8587 20.2931 18.8529 17.6329 16.6175 15.8409 13.8028 13.2089 11.631 Phase III Filtered 23.0712 21.8202 20.3187 18.9254 17.6648 16.6313 15.8391 13.8028 13.1488 11.6363

It can be seen from the above table that the method of the present invention comprises two stages of processing data, and the data of the first stage is substantially similar to the WFMFilter, but after two stages of processing, the data and effects can exceed the WFM Filter by 3 to 4 dB, and even When the pulse noise height is increased to 200 (the Gaussian noise variation is 1), the data exceeds the wfm Filter by more than 5dB, indicating that the AMFG Filter is for filtering Gaussian pulse noise compared to WFM 18 1273514.

Filter has a better effect. In summary, the present invention uses a two-stage image filter. In the filter portion of the first stage, the first noise filtering operation is performed on the image contaminated by the noise first, because the noise ratio is greater than More than 3% of the images may have more than one pixel of noise after the first stage of filtering, in order to make the image after filtering the high frequency noise closer to the original image, so The Threshdded image is obtained by thresholding the image after the first stage of filtering the noise, and then the Thresholded image is used as the noise judgment of the second stage filter for the image after the noise has been filtered out. After processing again, it has been experimentally proven that the secondary processing of AMFG Filter* can greatly increase the image ancestor and PSNR value, so that the image noise filter can make the image clearer. In addition, as described in the foregoing method, the time and space complexity of the proposed image filtering, criticalization 'edge grabbing, and image segmentation can be respectively as shown in the following table, and the complexity is lower than the existing methods.

AMFG Filter algorithm space complexity

Phase I Filter #Image side length, Z image Histogram, ^^7---- Memory instantaneous maximum usage Ιΰϊ--*--_ Final storage memory usage 3Ν2 ' ~------- Phase Π Filter image side length "^55^ each side length. One ~_------- Memory ^ Instant Usage Ύ2 ^ ----------_> Final Storage Memory Usage 3N2 ~' —~---- One 丨 — 19 1273514

Phase I Filter Equals Number of Additions Number of Multiplications Number of Divisions Number of Comparisons

Floor number overall Phase: II Filter equal number """ comparison times AMFG Filter algorithm time complexity image side length; sampling square side length; preset valley point; heart = image Histogram sampling number, cy= The number of gray scales in the number of sampling segments; £=the total grayscale number of the image Histogram (c"xcp); h=the fuzzy mean number (Thresholds). 2cn+3N2+4((N2^k2xn2)HN2>< K2))^6k2+7(N2+cn)^5 = 0(N2xk2xn2+L) -N2+2(N\k2^l))-Jr3cn^4N^7(N2xk2><nJ-i- k2) ^N2xk2^\〇Cn+\ = 0(N2xh xn2+L) 〇(^χ々)+2(#χ 怂)+3(妒>< 幻x Y)+2 = 0(tV2x 幻~N^ 2(N2^k2^n2^ Cn)+(N2xk2) +1 = Q(tV2x^2+Z)~ % Move (^ with Y+蝌(8)+l)(A^cg+(iV-11))+37^ = tear 2 ' xL) " cn+l = 0(cn) -~~' 0(N2 xk2xn2 Jrk2xL) " ~ ^ iV=image side length; π sample square side length. '~ ^ [2("2 -1) + φν a 2)2 ~[5〇2 a l) + 3j(iV-2)2 (n2-l)(N-2)2 =0(π2χΝ2) Accumulate Number of times h~h2 - 1)(#2)2=〇(Aa/·2) 2(iV - 2)2 =〇([2(^-D + 2j(iV-2)2 -0( η2χΝ2)

In terms of image analysis, the criticality, edge grabbing and image segmentation effects are reduced with such low complexity. (IV) The analysis of uncontaminated images of other green riders of the US Treasury breaks through the existing technology for the analysis of highly polluted images. territory. Such contributions are of great value to industries such as medical imaging cameras, image compression, and multi-hull. Therefore, the method is not only industrially viable and has high ambiguity value, especially the image τη after multi-value segmentation in the first segment, and can also be replaced by images generated by other multi-value segmentation algorithms, and is used as the second-stage processing method. The input is expanded to broadly apply. Therefore, the embodiments of the present invention have indeed achieved the intended purpose and efficacy. 'There is no disclosure of the _ 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者 者After the conclusion, and granting a patent, I am deeply impressed. [Simple description of the diagram] The first-image image sampling window is horizontally scanned and moved. The second diagram uses fuzzy inference to perform a schematic of image processing. The first figure is a flowchart of image filter processing in the -pg segment of the present invention. The fourth figure is a flow chart of the image filtering n process of the second stage of the present invention.

Figure 5 is a schematic diagram of the two-stage image processing of the present invention. Figure 6 (a) Baboon original image (b) Baboon doped with 70% height 150 pulsed noise image. (c) Baboon MEDIAN Filters filter out images after noise. (d) Baboon WFM Filter filters out images after noise. (e) The first stage of Baboon AMFGFilter filters out the image after the noise. (f) The second stage of Baboon AMFGFilter filters out the image after the noise. Figure 7 (a) Baboon WFM Filter multi-valued image. (b) Image of the first stage of Baboon AMFG Filter multi-valued. (c) Baboon AMFG Filter Phase 2 Multi-Valued (d) Image captured by the Baboon WFM Filter edge. (e) Image of the first stage edge capture of Baboon AMFG Filter. (f) Image of the second stage edge capture of BaboonAMFGFilter. Figure 8 (a) Lenna original image (b) Lenna is doped with a 70% height 150 pulsed noise image. 21 1273514 (c) Lenna MEDIAN Filters filter out images after noise. (d) Lenna WFM Filter filters out images after noise. (e) The first stage of the Lenna AMFGFilter filter out the image after the noise. (f) The second stage of the Lenna AMFGFilter filter out the image after the noise. Figure IX (a) Lenna WFMHlter multi-valued image. (b) Lenna AMFGFilter first-stage multi-valued image. (c) Lenna AMFGFilter Phase II Multi-Valued

(d) Image captured by the edge of the Lenna WFMFilter. (e) Image of the first stage edge capture of Lenna AMFGFilter. (f) Image of the second stage edge capture of LennaAMFGFilter. Figure 10 (a) Original Peppers image (b) Peppers doped with 70% height 150 pulsed noise images. (c) Peppers MEDIAN Filters filter out images after noise. (d) Peppers WFMFilter filters out images after noise. (e) P-printed AMFGFilter The first stage of filtering out the image after the noise. (f) P-printed AMFGFilter The second stage filters out the image after the noise. Figure 11 (a) Peppers WFMFilter multi-valued image. (b) Images of the first stage of the Peppers AMFGFilter multi-valued. (c) Peppers AMFGFilter Phase 2 Multi-Valued (d) Images captured by the Peppers WFM Filter edge. (e) Images of the first stage of the Peppers AMFGFilter capture. (f) Image of the second stage edge capture of Peppers AMFGFilter. twenty two

Claims (1)

1273514 Milk Year i February 4th Repair (for the original 10, the scope of the patent application · · 1 · A high-contamination image analysis method using fuzzy segmentation, which includes a two-stage filter, where: The first stage of the image filter Statistically generating a histogram on the contaminated image, using the histogram of the noise image to automatically generate the LR parameter left spread of the Membership Functions (Left spread) α, mean (mean) m, and right spread (Right spread) β, used to represent its LR fuzzy set (α, m, /5) lr, and then use the automatically obtained fuzzy set and membership function to perform fuzzy according to fuzzy theory The signal filtering and generating the first stage of the critical image (Thresholded Image); in the second stage of the image ship, the image obtained from the first stage is resampled for horizontal scanning and converted into a less complex image. The action of the critical image generated by the previous part will be compared with the surrounding values of the sampled towel, whether it is the first stage of the city, and the part of the limb sampling window is not the impurity point. As a basis for filling the center point. 2. The high-contamination image analysis method for the separation of the job described in the first paragraph of the patent scope is as follows: (1) Substituting the image into the matrix and calculating The histogram of the matrix; (2) Dividing the histogram into several paragraphs and averaging each paragraph; (3) If the difference between the average values of each paragraph changes from a negative value to a positive value, it is preset as Gu Hezhan; 25 1273514 • '(4) In the two __ set up the maximum value of the grain, the gray level at the maximum is used as the peak, and then the minimum of the histogram between two adjacent ♦ points, the minimum The gray level at the value is used as the last correct valley point; (5) The minimum and the highest gray level are used as the two valley points, and the use of -^^^^^^^^^^^^^^^^^^^ Lr fuzzy set (1 m, yS) LR mean parameter m, the distance from the peak point to the left adjacent valley point is the left distribution parameter cc, and the distance from the peak point to the right adjacent valley point is the right distribution parameter β; (6) sampling Each generation of _ minutes, m, and the degree of membership of the window can be obtained for each pixel in the sampling window. The weighting of the cut gamma and the weighted average of the sampling window for the segmentation; (7) For the any-sampling window, calculate the domain with the smallest difference between the best energy estimate and the weighted average of each segment of the fuzzy segmentation The average value is taken as the round; thus, the image after the first stage of filtering out the noise is obtained. 3 · As in the patent application scope item 2, the high-contamination image analysis method for secret separation, in the first stage of the image filter in the first stage, we only take out the histogram at the maximum gray level value -5 and the minimum The grayscale value is between +5, so that most of the burst noise Impulse Noise in the image is filtered out. 4 · If the high-contamination scene using fuzzy segmentation is described in item 2 of the scope of the patent application, the gray-scale value in the image after the first-stage filtering is updated to generate a critical image (threshold image) ), the method is that all gray scale values 26 1273514 falling between two adjacent valley points (including smaller valley points) in the image after the first stage of filtering out the noise are replaced by their relative peaks After that, it is the first stage critical image. 5. If you apply for a patent range! In the high-contamination image analysis using fuzzy segmentation, the parameters α, m, and β are substituted into the fuzzy number to obtain the k-group membership function value ^ y to calculate the maximum possible estimated value of the girl-like wealth, multiplied by the Wei window value. The upper one is estimated to be worth the weighted value, and the weighted value divided by the membership function is the mean value. 6. For example, if the towel material _ surrounding the second stipulations of the high-staining image analysis method of Guan Ling cut, the difference between the weighted average value and the maximum measurable value is the most ^ then the output is extended to - towel, and repeated This process is completed until the image processing is completed, to complete the first stage image filter for noise filtering. 7 · If the application is specifically _ 丨 所 使 赖 赖 赖 高 高 高 高 赖 赖 赖 其 其 其 其 其 其 其 其 其 其 其 影像 影像 影像 影像 影像 影像 影像 影像 影像 影像 影像 影像 影像 影像 影像 影像 影像The image is placed in an NxN matrix of x(I,J); (2) The sampling medulla xd (I+i) is used for the TH(I,J) matrix, and the horizontal scanning is performed; (3) Kt compares the center point of the sampling window with each point around it;) If: >, the Kt points have the same gray level value as the frame point, then the center point is judged as a noise point; (5) Take out the sampling window Deduct the value of the non-noise point outside the center point; and do the intermediate value calculation; (6) Substituting the center point determined as the noise point into the value after the intermediate value operation, each taking 27 1273514 • The sample window is reduced by the above The reading is _ second-order 郷 _ gift image; so after two stages of _ noise processing can greatly increase the image SNR and PSNR value, so that the image noise router can make the image clearer. —8. As in (4) the second step of the fourth paragraph or the seventh item of the patent, the dyeable image analysis method, in the fourth step of the image filter in the second stage, may set the noise point. The threshold value is judged as Kt, and the judgment value of the noise point is κ. The center point of the sampling window is divided into different navigational values, and the t-value _ miscellaneous value κ is accumulated. When the φ value is not the same, the judgment value κ is unchanged, when K< When Kt, the center point is determined to be a noise point, and when K^Kt, it is determined that the center point is not a noise point. 9. If you apply for the sixth or seventh remuneration of the special fiber, the second stage image noise filtering image generated by the image is analyzed, and then the first stage filter and impurity removal are performed. The grayscale value of the image towel is updated to generate the critical image ahreshoMed lmage), which is the result of the first stage filtering out the noise in the image between the two adjacent valleys (including the smaller valley). The grayscale value is replaced by the grayscale value of its relative peak, and the final critical image can be obtained. # 1〇·If the high-contamination image analysis method using fuzzy segmentation described in item 7 of the patent application is applied, the final critical image is shot, and the last critical image is sampled. The center point in the sample window is left and left. Above, the upper right four adjacent points are set to the maximum gray level value if f 'if any point is different', otherwise set to 0, all the edges of the window can be obtained after the edge detection image (Edge Detected Image), the edge texture It is white. 1 1 · The high-pollution image analysis method using fuzzy segmentation as described in item 7 of the patent application's method of obtaining the image of the towel, the center of the sampling window 28 1273514 points and the left, upper left, upper, upper right Compare the four adjacent points. If any point is different, set it to 0. Otherwise, set it to the maximum gray level value. After all the sampling windows are completed, you can get the edge detection image of the black edge (Edge Detected Image). The edge texture is black. . 12. The high-pollution image analysis method using fuzzy segmentation as described in claim 1, wherein the horizontal scanning method may be from top left to bottom right 'upper right to lower left, lower left to upper right, and lower right to upper left. 13. The high-pollution image analysis method using fuzzy segmentation as described in the first paragraph of the patent application, wherein the horizontal scanning method used can also be replaced by a vertical method. 14. The high-pollution image analysis method using fuzzy segmentation as described in item 1 of the patent application section, wherein the sampling window in the image filter can be contaminated by noise of the image noise image in addition to 3> To the extent that the sampling window is zoomed in, such as the 5x5, 7x7, 5x7, and 7x5 sampling windows, the larger noise can be removed, allowing the processed image to be closer to the original image. 15. The high-contamination image analysis method using fuzzy segmentation as described in item 1 of the scope of the patent application, the reference value of the histogram segmentation is given according to the well-known Nike's KNyquisi Rate sampling theorem. Formula to obtain the number of sampling segments Cn. 1 6 · The high-contamination-image analysis method using the fuzzy segmentation as described in the scope of the patent application, the multi-valued fractal image TH of the first stage can also be generated by other multi-valued algorithms. The image f generation, (4) as the second stage of the processing method of the round. 17. The second stage of processing can be implemented multiple times using the high-polluting knife-analysis method using fuzzy segmentation as described in item i of the patent application scope. It is Phase III, 29 1273514 • Its PSNR and SNR will not change significantly. .