Sen Orers for Reprints to reprints@benthamscience.ae 594 The Open Cybernetics & Systemics Journal, 205, 9, 594-600 Open Access Aaptive Image Enhancement Algorithm with Complex Bacgroun Zhang Pai * epartment of Information Engineering, Tang Shan College, Tang Shan, 063000, P.R. China Abstract: How to capture an enhance the target from an image with complex texture an ege features is stuie in this paper. For the traitional image enhancement methos can not split ifferent ins of the textures an eges of an image efficiently, a novel aaptive enhancement algorithm base on nonsubsample contourlet transform (NSCT) is projecte. Firstly, by using the irect spatial multiplication of NSCT coefficients of each same irectional subban at ajacent highpass scales, the inter-scale spaces are constructe an the optimal geometrical irections of each matrix region are calculate in the propose spaces. Experimental results show that, in NSCT omain, the geometrical irection information which locates in the inter-scale spaces can represent the ifferent etail features exactly of an image. Therefore, the optimal geometrical irection information as a rule to etermine what in of ege an texture features woul be capture an enhance is employe. After comparing with existing enhancement metho, the propose approach achieves better results, not only clearly ientifies target eges an textures from complex bacgroun, but also enhances them with better visualization. Keywors: Complex bacgroun, Image enhancement, NSCT, Optimal geometrical irection information.. INTROUCTION Image enhancement is to improve the interpretability or perception of information in images for human viewers, or to provie better input for other automate image processing techniques. In this paper, a type of image is iscusse whose targets possess complex eges an high contrast texture features while the bacgroun also has complex eges but lower contrast texture features. The aim is to capture an enhance the targets accurately from the images. Most of traitional enhancement methos [-3] can enhance wea eges an features in an image while restraining the noise an eeping the strong eges, but they are invali in ientifying an enhancing the eges or textures from objective region in an image with complex bacgroun. To overcome the limitation of existing enhancement methos an mae full use of the test image features, we come up with a novel aaptive enhancement algorithm base on nonsubsample contourlet transform (NSCT) [4]. As a shift-invariant version of the contourlet transform [5], NSCT provies an efficient irectional multiresolution image representation. In NSCT omain, each of highpass irectional subbans shows a high egree of irectionality an anisotropy. Taing into account the coefficients epenencies across scale in multiresolution ecomposition an the features of test images, we use the irect spatial multiplication of NSCT coefficients of each same irectional subban at ajacent scales to construct the inter-scale spaces an implemente two concrete steps in the constructe interscale spaces to istinguish target s eges an textures from *Aress corresponence to this author at the epartment of Information Engineering, Tang Shan College, Tang Shan, 063000, P.R. China; Tel: 3933958293; E-mail: rainrop_pai@qq.com 874-0X/5 bacgrouns. Firstly, base on the iea of banelet transform [6], we compute the optimal geometrical information of sub-squares (of matrix region) in constructe inter-scale spaces. This geometrical information can create a full istinction between high contrast ege an texture features an lower contrast ege an texture features of the source image. Then, in terms of the geometrical irections, an aaptive enhancement metho was employe to preserve an enhance the goals, while weaening the bacgroun. In this paper, the motivation is to project a new enhancement algorithm which has ability in capturing an enhancing the target with high contrast features in comparison with the bacgroun of source images. Experimental results show the availability of our approach. 2. PRINCIPLE OF NSCT AN THE CONSTRUCTION OF THE OPTIMAL GEOMETRICAL INFORMATION IN ITS INTER-SCALE SPACES 2.. The Principle of NSCT In the foremost contourlet transform, ownsamplers an upsamplers are presente in both the laplacian pyrami (LP) an the irectional filter bans (FB). Thus, it is not shiftinvariant, which causes pseuo-gibbs phenomenon aroun singularities. NSCT is an improve form of contourlet transform. In contrast with contourlet transform, nonsubsample laplacian pyrami (NLP) structure an nonsubsample irectional filter bans (NFB) are employe in NSCT. The NLP structure is achieve by using two-channel nonsubsample 2 filter bans. The NFB is achieve by switching off the ownsamplers/upsamplers in each two- 205 Bentham Open
Aaptive Image Enhancement Algorithm with Complex Bacgroun The Open Cybernetics & Systemics Journal, 205, Volume 9 595 channel filter ban in the FB tree structure an upsampling the filters accoringly. As a result, NSCT is shift-invariant an leas to better frequency selectivity an regularity. Fig. () shows the ecomposition framewor of NSCT. Input image NSCT is shift-invariant such that each coefficient of the transform subbans correspons to that of the original image in the same location an the N _ Corr of each ajacent scales has the same property. On comparing with the coefficients of subbans in each single scale, N _ Corr over two ajacent scales sharpens an enhances major eges while suppressing small sharp features. Fig. (2) as an illustration shows the effectiveness of N _ Corr in representing the locations of eges an other significant features of images. Banpass irectional subbans Lowpass subban Fig. (): ecomposition framewor of NSCT. In contrast with contourlet transform, nonsubsample laplacian pyrami(nlp) structure an nonsubsample irectional filter bans (NFB) are employe in NSCT. The NLP structure is achieve by using two-channel nonsubsample 2 filter bans. The NFB is achieve by switching off the ownsamplers/upsamplers in each twochannel filter ban in the FB tree structure an upsampling the filters accoringly. As a result, NSCT is shift-invariant an leas to better frequency selectivity an regularity. Fig. () shows the ecomposition framewor of NSCT. 2.2. The Construction of the Optimal Geometrical Information in its Inter-Scale Spaces In wavelet omain, the irect correlation of wavelet coefficients was use by Xu an co-worers to selectively enoise images in an ege-oriente manner [7]. It has been observe that by multiplying the ajacent scales eges coul be sharpen while iluting noise. This paper proposes an NSCT base ege etection scheme by scale multiplication. The NSCT coefficients coming from same irectional subbans at two ajacent scales are multiplie as a prouct function. We efine the function as follows: N _ Corr = N! N () l l + Here, N _ Corr represents the prouct ata of NSCT coefficients in the irection at l an l + scales. Nl an N + are the NSCT coefficients of irection at ajacent l l an l+scales, respectively, where l ( l [, L ]! ", L is the total number of scales) is the number of scales involve in!,, is the total the irect multiplication an ( [ ] number of irections) is the number of irections in corresponing scales. Fig. (2): An illustration of NSCT coefficients in the same irection(where ) at several single scales(where ) an coefficient proucts at corresponing ajacent scales of an horse image. From top left to bottom right: Original image; N _ Corr ; N _ Corr. 2,3 N ; 2 N ; 3 N ; From Fig. (2), we can observe that, in each coefficient space, non-zero ata represent the singularities of source image. In N l spaces, too many ege an texture features are presente that can not istinguish them in an efficient way. spaces, only significant eges an textures In l, l are preserve while small eges an regular parts of source image are eliminate. Therefore, the geometrical information spaces is more effective an tractable. in l, l Next, we nee to exploit an efficient approach to istinguish ifferent ege an texture features in each space. Firstly, we woul lie to introuce a l, l brief review of the banelet transform. The banelet approximation scheme, introuce in [6], taes avantage of geometric image regularity by removing the reunancy of a warpe wavelet transform by performing a baneletization which is efine as reorering the 2 wavelet coefficients an then performing a wavelet transform. The classical tensor wavelet transform of an image is the ecomposition of the latter on an orthogonal basis forme by the translation an ilation of three mother wavelets!,!,! for the horizontal, vertical an iagonal { } H V irections. Once the wavelet transform is foun, the quatree is compute by iviing the image into yaic squares with 4
596 The Open Cybernetics & Systemics Journal, 205, Volume 9 Zhang Pai variable sizes (refer to [6] for more information on computing the quatree). For each square in the quatree the optimal geometrical irection is compute by the minimization of a lagrangian (refer also to [6]). Then, a projection of the wavelet coefficients along the optimal irection is performe [6]. Finally, a iscrete wavelet transform is carrie on the projecte coefficients. For each highpass scale of the wavelet transform, the quatree an a zoom on the orientation of the linear flow on each yaic square are require. Notice that the quatree segmentation performs very well. Base on the iea of banelet transform, we compute the optimal geometrical irection of each square with a fixe size in the propose N _ Corr spaces. The best irection is the one that leas to the best compresse representation after the baneletization. If there is no preferential orientation in the square S, then it is better not to perform any baneletization. In this case, we set irection is equal to NULL in the present square S. Fig. (3) as an illustration shows the optimal geometrical information of each square S in one of the N _ Corr spaces of the part of horse image. For the square whose irection is not equal to NULL, we now perform the iscrete reorering of the sampling location. This is one by projecting the sampling location along geometrical irection an sorting the resulting points from left to right. The obtaine numbering of the sampling points efines a iscrete signal f, an then performs a iscrete wavelet transform of f. We enote by { b } the coefficients of the wavelet transform of f. From { b }, one can see that the high wavelet coefficients corresponing to most of ege information are gathere, an this property provies convenience in the following processing. 3. IMAGE ENHANCEMENT ALGORITHM 3.. The Effectiveness of Image Eges Representation in the Inter-Scale Space When encountering a type of images whose targets are possesse of complex strong eges an high contrast texture features while the bacgroun also has complex strong eges but lower contrast texture features, we are incline to capture an enhance the target from complex bacgroun. Traitional enhancement algorithms lose efficacy in solving the problem. To mae full use of the own features of this type of images while processing them has become significant. Refer to [7] pointe out that the steep regions of source image yiel bigger coefficient amplitues in frequency omain while the smooth region is corresponing with small coefficient amplitues. In this section, we woul lie to explore the properties of ifferent ins of ege an texture features representation in propose space. The horse image is still taen as a test image ecompose by NSCT in several scales. Fig (4) shows each inter-space constructe by corresponing N _ Corr (where l = 3, l + = 4 an the total irection num is set as four which is the same in two ajacent scales). Fig. (4): An illustration of inter-scale spaces of all irections. From 2 left to right: Original image; N _ Corr ; N _ Corr ; N _ Corr ; 3 N _ Corr. 4 Fig. (3): The optimal geometrical information of each square S in one of the N _ Corr spaces of the part of horse image. From Fig. (4), it can be observe that the sie face of horse has high contrast textures but not so strong as textures of the front face are transforme into wea eges in the spaces, while the textures of the front face are l, l transforme into strong eges an the textures of the nose are suppresse in the N _ Corr spaces. Hence, it can be conclue that the wea eges in N _ Corr space are the representation of texture features with high contrast in source image. Now, to orientate wea eges accurately in N _ Corr space becomes an important issue. In NSCT omain, the
Aaptive Image Enhancement Algorithm with Complex Bacgroun The Open Cybernetics & Systemics Journal, 205, Volume 9 597 strong eges correspon to those pixels with big-value coefficients in all subbans. The wea eges correspon to those pixels with big-value coefficients in some irectional subbans but small-value coefficients in other irectional subbans within the same scale. The noises correspon to those pixels with small-value coefficients in all subbans. (refer also to [8]).This observation is also effective in spaces. epening on the geometrical irection l, l information of each fixe square region in N _ Corr spaces, we exploit a new approach by aing up to the number of geometrical irections which are unequal to NULL to etermine the class of eges in the present fixe square region. Then, repeat the previous steps to etect other sub-regions. The etaile escription will be taen in the next section. 3.2. The Implement Steps of the Enhancement Algorithm In this section, the implement steps of propose enhancement algorithm can be ivie into four ifferent stages with reference to Fig. (5). Stage I ) Mae the source image via NSCT an illustrate it. The mapping is assume to be of the form: 0 x! N, N2,! NL, xl. Where, Nl is the etail image at { } level l an xl is the approximation at the coarsest level L. Then, N, l =, 2! L is further separate into subbans l 2 accoring to its orientation, namely, Nl = { Nl, Nl,!, Nl } representing irections. Here, the number of irectional subbans at all high-frequency levels is the same with each other, which is for preparation of constructing the inter-scale coefficient prouct spaces in the same irectional subban across ajacent levels. 2) By using the irect spatial multiplication of NSCT coefficients of each same irectional subban at ajacent scales to construct the inter-scale spaces. N _ Corr represents the prouct ata of NSCT coefficients in irection at l an l!, L ", L is l + scales, where l ( [ ] the total number of scales) is the number of scales involve!,, is the total in the irect multiplication an ( [ ] number of irections) is the number of irections in corresponing scales. This is efine as previous section. 3) ivie N _ Corr into subregions (of 4! 4 matrix region) which can be enote as N _ Corr, N _ Corr 2,!, N _ Corr P. P is { } the total number of subregions. Then, we perform banazation to calculate the optimal geometrical irections N _ Corr p! + p as representation of of an efine l, l irection information of each p square region, especially,! p = Inf represents non geometrical information in N _ Corrl, l + p region. Here, we efine a new variable S + to ivie l, l l, l state as follows: S! + into two categories, which is "! # Inf = % & 0 else where l![, L " ],![, ], p [, P] Stage II!. ) In term of the geometric information features pixel by pixel of the NSCT coefficients emonstrate in section 3., we observe that in the same inter-scales, the wea eges probably exist in the sub-region whose! = Inf at a part of irectional subbans while at others,, the strong eges correspon to l, l p Inf (2)! l l + p " Inf,! + " through the whole irectional subbans, the small eges an smooth regions correspon to! " Inf at most of irectional subbans. These properties prompt us the total number of path regions whose! " Inf at all irectional subbans in present inter-scale can be use as a rule to classify the eges. Equation (3) is a statement of the concrete rule.! small ege 0 " ' S < # = 2 ege _ style = % wea ege # " ' S " # 2 = strong ege ' S = & = The threshols of etermining the ege styles are aopte by the test of numerous experiments. We further investigate to generate a map compose of 0 or as the representation of the ifferent type of eges. map (3)! wea ege = " (4) # 0 else 2) Here are two confusing issues: Firstly, owing to the complexity an irregularity of bacgroun in source images, wea eges may also exist in the region of N _ Corr spaces which correspon to the source bacgroun regions. This will cause ba accuracy in capturing the target s regions. Seconly, the strong eges are both in the target s an the bacgroun s region of source images, an how to ientify strong eges of the target region is another issue we nee to solve. To settle these ifficulties, we nee to consier the features of source images once again. Firstly, in source images, owing to the fact that texture features of bacgroun often possess lower contrast an not so steep, their corresponing sub-regions with wea eges in inter-scale woul be scattere an inepenent. Seconly, the target s strong eges are along with its texture features to compose an integral goal.
598 The Open Cybernetics & Systemics Journal, 205, Volume 9 Zhang Pai Hence, we affirm that the strong ege regions which surroun the wea ege regions in each inter-scale space are the corresponing eges belonging to the target in source image. In terms of these two analyses, we prouce an operator! " which consists of 3! 3 ernels of the form W = # an # #% & use it to operate the map matrix (formula 5). The function of this operator is to reset the centre element value, which can coincie with most of the neighbors aroun it in the present 3! 3 region. This implementation can remove the single region with singularity in the inter-scale spaces, meanwhile preserving the objective strong ege regions. The following formula is a statement of the concrete rule of the operator. map # if ( map( p + i, p + j) 6 % i! " = & %' 0 [ [,] j! " else The element values of new map matrix can represent the belongings of corresponing sub-regions exactly in interscale spaces. Now, we achieve in capturing the target s region in the source. In the next stage, we woul lie to exploit an effective approach to enhance the objective region. Stage III ) The mapl, l +,!, l [, L ]![, ], p [, P] (5)! ", an! are use as the information to operate NSCT corresponing irectional subbans in previous l!, L ". For the finest scale ( l = L ) represents l scale [ ] most of the strong eges feature, we are not taing it into account here. In orer to highlight the objective region, an operation is performe by enhancing the target meanwhile weaening the bacgroun. At each highpass irectional map p! + p as subban in NSCT omain, by using, l, l the rule, the corresponing sub-regions can be etermine whether to enhance or not. The concrete rule is shown as follows: [, ]; &[, ],!" enhance mapl l + p = sub # regionl =, "% weaen mapl, l + = 0 p & P 2) To provie an effective way in enhancing the objective region, we propose a new enhancement metho name Bb (has been mentione in algorithm by consiering { } previous section) as each sub-region s ata. The B- algorithm is implemente as follows: firstly, the threshol of each{ b } is calculate respectively in formula (7). Then, the coefficients greater than the threshol are regare as the high coefficients which represent the eges information in (6) the { b }, we enhance them with a suitable weight (here, it is set to 3). The corresponing formula is shown as follows: meian( b i, j ) T =, b( i, j)!{ b } 0.6745 (7) "# b( i, j)! 3 b( i, j) > T b( i, j) = %, b( i, j) { b } #& b( i, j) b( i, j) < T (8) here, b( i, j ) represents each coefficient of { } Stage IV b. Perform inverse wavelet transform on each { b }. Then, by using the processe coefficients reconstruct the enhance image via inverse NSCT. Source image Forwar NSCT Construct the inter-scale space Compute the optimal geometrical irections B- enhancement algorithm Inverse NSCT Enhance image Fig. (5): Schematic iagram of propose fusion algorithm. Intelligent setch technology is base on the stuy of esign thining, setches behavior, its characteristics, the paper Setch esign technology an computer support, each of which has its own strengths an weanesses [8]. It will be a very useful wor if their avantages are combine together. Especially, now the CA technology which orients the etaile esign stage has evelope perfectly. It can guarantee the support an assistance of the computer for the entire esign process, consistently unifying the entire esign process. Moreover, the outstaning characteristics of the computer itself are boun to raise the level of efficiency an esign at the early esign stage. Especially, it can enhance the setch esign innovation performance through smart innovation an technology. The ey to intelligent setch technology is in-epth unerstaning of the raft plan to the behavior. Setching behavior an esign thining are closely relate. The setches capture inherent human behavior thining outsie or concrete which is a necessary extension of the visual image of the esigner s perception. Intelligent setch technology for computer support, accoring to the characteristics of the computer itself, can be roughly ivie into four categories: Natural interaction behavior smart setch; Setch of expression an freeom of freehan setch input; esign intent capture an setch recognition; an setch-base geometric moeling. Fig. () shows the
Aaptive Image Enhancement Algorithm with Complex Bacgroun The Open Cybernetics & Systemics Journal, 205, Volume 9 599 computer support han-painte ceramic pattern setches technology framewor. 4. EXPERIMENTAL RESULTS In this section, two typical images with complex ege an texture features are taen as the experimental images. Our aim is to capture an enhance the target from bacgroun accurately. We compare the enhancement results by the propose algorithm with those by the B- algorithm in wavelet transform an enhancement algorithm in NSCT omain. In the first experiment, the input is a butterfly in the littery grass as shown in Fig. (6). From Fig. (7), we can observe that the propose algorithm removes most of the ege an texture features belonging to the bacgroun an also performs excellent result in enhancing the objective region in the test image when compare with other two liste methos. Fig. (7): Cactus image enhancement results. From top to bottom: Original image; enhance image base B- in wavelet omain; Enhance image base NSCT enhancement algorithm; Enhance image base the propose algorithm. Fig. (6) Butterfly image enhancement results. From top to bottom: Original image; enhance image base B- in wavelet omain; Enhance image base NSCT enhancement algorithm; Enhance image base the propose algorithm. From Fig. (6), it can be observe that, in the enhance image of B- enhancement algorithm in wavelet omain metho, some of the eges an textures from bacgroun are preserve an enhance unexpectely, which causes the final experimental result to be unsatisfactory in visual observation. The reason is that, the wavelets lac the important feature of irectionality an hence, they are not efficient in retaining textures an fine etails of images in wavelet omain. NSCT enhancement algorithm oes better jobs in enhancing the wea eges in the textures an istinguishing wea eges from noises, but losing efficacy in capturing the objective region an hence its enhance result is not excellent in visual appearance either. Our experiments show that the propose approach outperforms other enhancement methos in visual effect. Fig. (7), as a escription of a coccinella septempunctata in the cereus, will give another strong example. CONCLUSION In this paper, we propose a new algorithm in capturing an enhancing the objective region from an image with complex texture an ege features. The approach base on NSCT, by using the coefficient proucts of each inter-scale an the corresponing geometrical irection information in the propose spaces, we can etect the location of the target accurately from source image. Then, we project a B- algorithm to enhance the target. The simulation results inicate that the propose metho is visually superior to others in preserving an enhancing eges an textures of target. CONFLICT OF INTEREST The author confirms that this article content has no conflict of interest. ACKNOWLEGEMENTS eclare none. REFERENCES [] A. Laine, J. Fan, an W. Yang, Wavelets for contrast enhancement of igital mammography, IEEE Engineering in Meicine an Biology, pp. 536-550, 995.
600 The Open Cybernetics & Systemics Journal, 205, Volume 9 Zhang Pai [2] S. ippel, M. Stahl, R. Wiemer, an T. Blaffert, Multiscale contrast enhancement for raiographies: Laplacian pyrami versus fast wavelet transform, IEEE Transactions in Meical Imaging, vol. 2, no. 4, pp. 343-353, 2002. [3] K. V. Vele, Multi-scale color image enhancement, Proceeings of IEEE International Conference on Image Processing, vol. 3, pp. 584-587, 999. [4] A. L. Cunha, J. Zhou, an M. N. o, The nonsubsample contourlet transform: theory, esign an applications, IEEE Transactions on Image Processing, vol. 5, pp. 3089-30, 2006. [5] M. N. o an M. Vetterli, The contourlet transform: An efficient irectional multiresolution image representation, IEEE Transactions on Image Processing, vol. 4, pp. 209-206, 2005. [6] G. Peyr, an S. Mallat, Surface compression with geometric banelets, ACM Transactions on Graphics (TOG), vol. 24, no.3, pp. 60-608, 2005. [7] Y. Xu, J. B. Weaver,. M. Healy, an J. Lu, Wavelet transform omain filters: A spatially selective noise filtration technique, IEEE Transactions on Image Processing, vol. 3, pp. 747-758, 994. [8] J. Zhou, A. L. Cunha an M. N. o, Nonsubsample contourlet transform: Construction an application in enhancement, Proceeings on IEEE International Conference on Image, vol., pp. 469, 2005. Receive: September 6, 204 Revise: ecember 23, 204 Accepte: ecember 3, 204 Zhang Pai.; Licensee Bentham Open. This is an open access article license uner the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/4.0/) which permits unrestricte, non-commercial use, istribution an reprouction in any meium, provie the wor is properly cite.