Sie sind auf Seite 1von 6

Cognition Model Based Adaptive Multilevel Filter for Removal of Impulse Noise

Mrs. M.Brindha, M.E, Lecturer, "ept. of Information !echnolog#, $%N College of Engg., & !echnolog#, !irunelveli m'rindhait(gmail.com
AbstractImage denoising is an important image

A.Radha, B.E, II M.!ech, "ept. of Information !echnolog#, $%N College of Engg., & !echnolog# !irunelveli radhachocostic)(gmail.com
think about the randomness. &mong these filters, the famous ones are the median (M*+ filter (,-., ,/. . The# unconditionall# fulfill on each pi$el without considering whether the pi$el is 0bad1 or not., since the uncorrupted pi$els are altered, the# damage man# image details in the high noise levels! & novel effective filter based on the cloud model (CM for impulse noise removal is presented, called the CM filter. The e$perimental results show that, compared with the traditional filters, the CM filter has the better performance in image denoising.

processing task, both as a process itself, and as a component in other processes. The main properties of a good image denoising model are that it will remove noise while preserving edges. This paper presents a novel adaptive detail-pre-serving lter based on the cloud model (CM to remove impulse noise! CM is an uncertain cognitive model called the CM lter. "irst, an uncertaint#based detector identies the pi$els corrupted b# impulse noise. Then, a weighted multilevel arithmetic mean lter is applied to remove the noise candidates. Compared with the traditional lters, the CM lter makes a great improvement in image denoising. Index terms- Cloud model (CM , image denoising, impulse noise, median filter.

II. CM FIL!ER
A. Preprocessing
The gra# values of the pi$els are usuall# lower than in the other areas. The images restored b# the CM filter basicall# keep the same gra# levels with the original images. Then the impulse noise is added. In the model, the observed gra# level at location (i.2 is given b#,

I. IN!R*"+C!I*N
Image denoising is the recover# of a digital image that has been contaminated b# additive white %aussian noise. &mong the uncertainties involved in impulse noise, the randomness and the fu''iness are the two most important features. The randomness mainl# shows in two aspects, i.e., the pi$els are randoml# corrupted b# the noise and the noise pi$els are randoml# set to the ma$imum or minimum value. The fu''iness focuses on the pi$els with the e$treme values whether the# belong to the noise or not. (ot all of the pi$els, which are set to the e$treme values, will be the noise pi$els. In earl# denoising techni)ues, the filters onl#

is helpful to identif# the noise.


B. Uncertainity

Detector

The CM is a natural-language cognitive model with uncertaint#. It combines the fu''iness and the randomness, and forms an inter mapping between the )ualitative and )uantitative information.

Then the distribution of $ on 3 is called the cloud, and each $ is called a drop ,-4.. The cloud can be characteri'ed b# three parameters, i.e., the e$pected value *$, entrop# *n, and h#per entrop# 5e. *$ is the e$pectation of the cloud drops6 distribution in the domain *n is the uncertaint# measurement of the )ualitative concept, which is determined b# both the randomness and the fu''iness of the concept.
"ig. /. (a Calculated the cloud that represents the observed neighborhood. (b Cloud C (-74.:, ;;.<, and ;:.= represents the neighborhood in (a .

The contribution degrees and the certaint# degrees are usuall# lower than the others. It is how we can distinguish the noise pi$els from the uncorrupted ones, e.g., there is an observed neighborhood (the left s)uare region of "ig. /(a !
"ig. -. Cloud C (7, /, 7.8 .

7 and /44 are the noise pi$els and if cloud e$ists, which can represent the neighborhood. >et each pi$el be a cloud drop and input them into the backward CM generator C% --. The outputs of C%-- are the three parameters of cloud. Then, we input *$, *n, and 5e into the forward C%. "inall#, cloud comes out as the output of the C%.
T&?>* I C*@T&I(TA +*%@** B" *&C5 CID*>

The cloud emplo#s its three parameters to represent the )ualitative concept. "or e$ample, cloud *$ *n 5e is shown in "ig. -. The drops compose the cloud. 9hen the drops are approaching e$, the certaint# degrees and the contribution degrees of the drops are increasing. Therefore, in the cloud, the drop communities contribute to the concept with the different contribution degrees. &ccording to the normal cloud generator (C% , the certaint# degree of each drop is a probabilit# distribution rather than a fi$ed value. It means that the certaint# degree of each drop is a random value in a d#namic range. If 5e of the cloud is 7, then the certaint# degree of each drop will change to be a fi$ed value. The fi$ed value is the e$pectation value of the certaint# degree. In fact, the value is also the unbiased estimation for the average value of the certaint# degrees in the range. &ll the drops and their e$pectations of certaint# degrees can compose a curve, and the curve is the cloud e$pectation curve (C*C . "or e$ample, the red curve is the C*C of cloud (see "ig. - . Thus, the noise pi$els are usuall# distributed on the both sides of the cloud, and the uncorrupted pi$els are located near the central region of the Cloud.

Table I shows that the certaint# degrees of the noise pi$els are far less than that of the uncorrupted pi$els (the certaint# degree of each pi$el, which is calculated through the C*C .

C. Noise Removal
The impulse noise is removed from the image. Calculate *$ and *n. The filter replaces the noise pi$el b# using the weighted mean of the remaining pi$els, and their weights are the certaint# degrees of them. In the cloud, the certaint# degree of each drop is a random value.

D. Weighted Multilevel Arithmetic Mean lter

Algorithm I

Bnce the CM filter identifies a pi$el as a 0good1 one, the pi$el naturall# keeps its original value. Bnl# the corrupted candidates are replaced, which is the same to the traditional filters. 5owever, man# switching methods are two-stage filters. The# identif# the noise pi$els first and then use a noise map to record the information of the noise pi$els, such as the pi$el locations. "inall#, according to the map, the filters remove the noise pi$els one b# one. Thus, the# scan the noise image twice. Those filters not onl# increase the memor# spaces but also decrease the computational efficienc#. To overcome this drawback, the CM filter removes a pi$el immediatel# after the pi$el has been identified as a corrupted candidate. Therefore, in the CM filter, the noise detector and the post filter use the same windows. It means that the window si'e of the post filter is the one that is used b# the noise detection at the last time. "or e$ample, in a 8$8 window, the CM filter cannot identif# if a pi$el is 0good1 or 0bad.1 Then, the window si'e will be adaptivel# increased. 3ntil in the :$: window, the pi$el is identified as a corrupted candidate, and the CM filter removes the pi$el in the same :$: window immediatel#. Man# denoising methods are also switching median filters. It means that the filters tr# to identif# the noise pi$els and then replace the noise pi$el b# the median value.

Algorithm II

II. RE%+L!% AN" "I%C+%%I*N A. Con!iguration


Two commonl# tested 4-/$4-/ =-bit gra#scale images, >ena and ?ridge are selected in the simulations. The images are corrupted b# e)ual probabilit# 0salt1 (with value /44 and 0pepper1 (with value 7 noise. "or comparative purposes, the adaptive median (&M median-t#pe 5owever, in the cloud, the certaint# degree of each drop is a random value. Thus, to increase the computational efficienc# and the robust stabili'ation of the CM filter, the filter also uses the C*C to calculate the certaint# degree for each pi$el. regulari'ation noise detectors filter and ,-/., filter ,8., the detail-preserving the boundar# minimumEma$imum e$clusive mean (MM*M filter, the (&M-*C@

discriminative noise detection (?+(+ filter, and the fast median ("M filter ,-;. are also tested. These filters can remove the salt-and-pepper noise in the high noise levels. 5owever, when the noise level is higher than F7G, the other filters (,=., ,4. cannot remove the noise with good

image )ualities. "or an in-depth stud# in the denoising performance of the selected filters, the simulations are divided into multiple stages. "irst, the filters appl# on the noise images in a wide range of noise levels var#ing from -7G to =7G with increments of -7G. It focuses on two aspects, the accurac# of the noise detection and the )ualit# of the restored image. Therefore, onl# the filters that can restore the images without noise and distortion will enter the ne$t stage. Hecond, the filters that passed the first stage will be applied on the nois# images with the highest noise level (<7G . The e$periment aims to stud# the detailpreserving abilities of the filters when the images are affected b# a severe noise. "inall#, the CM filter with different values applies on the noise image in a wide range of the noise levels var#ing from -7G to <7G with increments of -7G. The main ob2ective is to characteri'e the robustness to the threshold parameter .The "M filter directl# regards all pi$els whose values are set7 and /44 in a fi$ed-si'e (8$8 detection window as the corrupted pi$els and then uses the median values or the left neighborhood values to replace the noise pi$els. The MM*M filter discards all pi$els whose values are e)ual or similar to the ma$imum or minimum values in 8$8 or 4$4 windows and then calculates the average value (&I% of the remaining pi$els or the four neighboring pi$els. "inall#, if (is the detected pi$el , then the detected pi$el is a noise pi$el, and it will be replaced b# &I%. The &M filter uses an adaptive si'e window to identif# the noise pi$els and then replaces the noise pi$els b# the median values. To detect the noise patches and filter out the noise, the ma$imum window si'e should be chosen such that it increases with the noise level. Therefore, set in all our simulations. The &M-*C@ filter combines the &M filter and a variation method. It uses the &M detector to identif# the noise pi$els and then replaces the noise pi$el b# using the variation method to preserve the image details. Thus, for the &M-*C@ filter, we also set in all our simulations. In addition, we choose as the edgepreserving function. 9e observe that, if, most of the noise is suppressed, but staircases appear. If, the fine details are not seriousl# distorted, but the noise cannot be full# suppressed. &nother parameter of the &M-*C@ filter is the pertinent choice factor. In tests, the &M-*C@ filter is ver#

robust with respect to. Thus, we fi$ in all our simulations. The ?+(+ filter uses two si'e windows (8$8 and /-$/to identif# the noise and then uses a noise map to recorder all noise pi$el addresses. "inall#, it uses an adaptive window to remove the noise. In order to avoid severe blurring of image details at high-noise-densit# cases, the ma$imum window si'e of the adaptive window is limited to :$: in all our simulations. In addition, to increase the computational efficienc# of the CM filter, the detection threshold is alwa#s - in the first-stage simulations, and to increase the )ualities of the restored images, changes to 8 in the second stage.

B.

Noise Detection Per!ormance

The denoising performances of the switching filters are usuall# higher than the standard median filter and its varieties, because the switching filters onl# remove the noise without altering the uncorrupted pi$els. Therefore, the noise detection pla#s a ke# role in image denoising. 5owever, with the noise level sharpl# increased, the noise patches will be formed. The pi$els in the noise patches are eas# to be identified as the 0good1 ones, which often results in detection errors. Thus, the accurac# of the noise detection can directl# influence the )ualities of the restored images.

C. Restoration Per!ormance
The restoration performances are )uantified b# the peak signal to-noise ratio (CH(@

It denotes the pi$el values of the restored image and the original image, respectivel#. 9hen the noise level is lower than F7G, the performance of the CM filter is similar to the ?+(+ filter and the &M-*C@ filter. 5owever, with the noise level sharpl# increased, man# noise patches will be formed. This fact causes man# detection errors and makes, the differences between each filter result more and clearer. The "M filter creates man# stripe regions because it often replaces the corrupted pi$el b# the left neighborhood pi$el.

image restored b# the &M-*C@ to decrease. +ifferent outcomes between two filters are represented b# difference images. The difference images are derived from the absolute value error images b# using both original and restored images after filtering.

@estoration results of different filters. (a Corrupted >ena image with =7G salt-and-pepper noise (F.;/ d? . (b Briginal image. (c CM filters (/=.FF d? . (d MM*M filter (/:.FF d? . (e &M filter (/;.=< d? . (f ?+(+ filter (/:.F: d? . (g "M filter (/8.7= d? . (h &M-*C@ filter (/:./8 d? . (a +ifference image of >ena using the CM filter at the =7G noise level. (b +ifference image of >ena using the &M-*C@ filter. (c +ifference image of ?ridge using the CM filter at the =7G noise level. (d +ifference image of ?ridge using the &M-*C@ filter.

&lthough having the same noise detector as the &M filter, the &M-*C@ filter makes a great progress in the post filtering. The &M filter cannot preserve the edges well at the high noise level ,see "igs. ;(e and 4(e ., because it is a switching median filter. To overcome this drawback, the &M-*C@ filter combines the &M filter and a variation method to preserve the image details. Thus, the )ualities of the images restored b# the &M-*C@ filter are usuall# better than the &M filters.

5owever, at the edges of the >ena hair, the color of "ig. (a is darker than that of "ig. (b , i.e., in these high activit# regions, the edge-preserving abilit# of the CM filter is better than that of the &M-*C@ filter.

@estoration results of different filters. (a >ena with the noise level of <7G (4.<7 d? . (b Briginal image. (c CM filters (/F.=4 d? . (d ?+(+ filters (/4.;4 d? .

@estoration results of different filters. (a Corrupted ?ridge image with =7G salt-and-pepper noise (F.// d? . (b Briginal image. (C CM filters (//.F8 d? . (d MM*M filters (/-.FF d? . (e &M filter (/7./- d? . (f ?+(+ filters (/-.=/ d? . (g "M filters (/7./; d? . (h &M-*C@ filter (/-.:7 d? .

>ocal restoration results (>ena of different filters. (a (oise image. (b Briginal image. (c CM filter. (d ?+(+ filter.

Therefore, the images restored b# the &M-*C@ filter appear to have man# speckles, particularl# in the regions of the >ena hair and the bottom right corner of ?ridge. In the images, these regions are the highest activit# regions, in which the gra# values of the pi$els are usuall# lower than in the other areas. It means that the differences between the pi$el and its neighborhood pi$els are smaller than the others. These two reasons reduce the sensitivit# of the detail-preserving method and cause the )ualities of the
>ocal restoration results (?ridge of different filters. (a (oise image. (b Briginal image. (c CM filters. (d ?+(+ filters. @estoration results of different filters. (a ?ridge with the noise level of <7G (4.:- d? . (b Briginal image. (c CM filter (/-.87 d? . (d ?+(+ filters (/7.-8 d? .

,8. 5. 5wang and @.&.5addad, 0&daptivemedian filtersL (ew algorithms and results,1 I""" #rans. Image Process., vol. ;, no. ;, pp. ;<<E47/, &pr. -<<4. ,;. *. &breu, M. >ightstone, H. K. Mitra, and K. &rakawa, 0& new efficient approach for the removal of impulse noise from highl# corrupted images,1 I""" #rans. Image Process., vol. 4, no. F, pp. -7-/E-7/4, Jun. -<<F. ,4. T. Chen, K.-K. Ma, and >.-5. Chen, 0Tri-state median filter for image denoising,1 I""" #rans. Image Process. , vol. =, no. -/, pp. -=8;E-=8=, +ec. -<<<. ,F. M. 9ang and +. Mhang, 0Crogressive switching median filter for the removal of impulse noise from highl# corrupted images,1 I""" #rans.Circuits $yst. II% Analog Digit. $ignal Process., vol. ;F, no. -, pp.:=E=7, Jan. -<<<. ,:.+. Mhang and M. 9ang, 0Impulse noise detection and removal using fu''# techni)ues,1 "lectron. &ett.% vol. 88, pp. 8:=E8:<, "eb. -<<:. ,=. T. Chen and 5. 9u, 0&daptive impulse detection using center weighted median filters,1 I""" $ignal Process. &ett., vol. =, no. -, pp. -E8, Jan. /77-. ,<. 5.->. *ng and K.-K.Ma, 0(oise adaptive soft-switching median filter,1I""" #rans. Image Process., vol. -7, no. /, pp. /;/E/4-, "eb. /77-. ,-7. H. Mhang and M. &. Karim, 0& new impulse detector for switchingmedian filters,1 I""" $ignal Process. &ett., vol. <, no. --, pp. 8F7E8F8,(ov. /77/. ,--. I. Crno2evic, I. Nenk, and . Trpovski, 0&dvanced impulse detection based on pi$el-wise M&+,1 I""" $ignal Process. &ett., vol. --, no. :, pp. 4=<E4</, Jul. /77;. ,-/.@. 5. Chan, C.-9. 5o, and M. (ikolova, 0Halt-andpepper noise removal b# median-t#pe noise detectors and detail-preserving regulari'ation,1 I""" #rans. Image Process., vol. -;, no. -7, pp. -;:<E-;=4, Bct. /774. ,-8.%. Cok, J.-C. >iu, and &. H. (air, 0Helective removal of impulse noise based on homogeneit# level information,1 I""" #rans. Image Process., vol. -/, no. -, pp. =4E</, Jan. /778. ,-;.K. H. Hrinivasan and +. *bene'er, 0& new fast and efficient decision based algorithm for removal of highdensit# impulse noises,1 I""" $ignal Process. &ett., vol. -;, no. 8, pp. -=<E-</, Mar. /77:. ,-4.+. A. >i, C. A. >iu, and 9. A. %an, 0& new cognitive modelL Cloud model,1 Int. '. Intell. $yst., vol. /;, no. 8, pp. 84:E8:4, Mar. /77<.

Images with the noise level of <4G restored b# the CM filter. (a >ena (/;.<: d? . (b ?ridge (/7.74 d? . (c Ceppers (/;.;4 d? . (d ?aboon (-<.7= d? .

III. C*NCL+%I*N & novel filter with uncertaint# for impulse noise removal has been proposed. It represents the uncertainties of the noise perfectl# b# using the CM, which is helpful in detecting and removing the noise. In addition, the proposed filter identifies the noise pi$el without needing to sort the pi$el gra# values, which immensel# increases the computational efficienc# in noise detection. The e$perimental results show the CM filter is the best one among the tested filters, compared with the traditional filters, the CM filter makes a great improvement and has the higher performances. *ven if the noise level closes to <4G, the te$ture, the details, and the edges of the images restored b# the CM filter are preserved with good visual effect. In sum, the CM filter is a moderatel# fast denoising filter with good detail preservation. &lthough the CM filter can onl# detect the fi$ed-valued impulse noise, and at the <4Gnoise level, the restored images have some blurring edges in some local areas. This paper presents a novel adaptive multilevel lter based on the cloud model (CM to remove impulse noise. &n uncertaint#-based detector identies the pi$els corrupted b# impulse noise. & weighted multilevel arithmetic mean lter is applied to remove the noise candidates. The e$perimental results show that, compared with the traditional lters, the CM lter makes a great improvement in image denoising. 5owever, we can further improve the CM filter b# using different noise detectors or restoration methods to solve it. REFERENCE% ,-. H.-J. Ko and H.-J. >ee, 0Center weighted median filters and their applications to image enhancement,1 I""" #rans. Circuits $yst., vol. 8=, no. <, pp. <=;E<<8, Hep. -<<-. ,/. >. Ain, @. Aang, M. %abbou2, and A. (euvo, 09eighted median filtersL & tutorial,1 I""" #rans. Circuits $yst. II% Analog Digit. $ignal Process., vol. ;8, no. 8, pp. -4:E-</, Mar. -<<F.

Das könnte Ihnen auch gefallen