personalcommunicationtechnologiesforsmartspaces...

ResearchArticlePersonal Communication Technologies for Smart SpacesDensity-Based Clustering for Content and Color AdaptiveTone Mapping

Maleeha Javed1 Hassan Dawood 1 Muhammad Murtaza Khan23 Ameen Banjar4

Riad Alharbey4 and Hussain Dawood 5

1Department of Software Engineering University of Engineering and Technology Taxila Pakistan2Department of Computer Science and Artificial Intelligence College of Computer Science and Engineering University of JeddahJeddah 21589 Saudi Arabia3School of Electrical Engineering and Computer Science (SEECS) National University of Sciences and Technology (NUST)Islamabad 44000 Pakistan4Department of Information Systems and Technology College of Computer Science and Engineering University of JeddahJeddah 21589 Saudi Arabia5Department of Computer and Network Engineering College of Computer Science and Engineering University of JeddahJeddah 21859 Saudi Arabia

Correspondence should be addressed to Hassan Dawood hasandawodyahoocom

Received 11 March 2020 Revised 25 June 2020 Accepted 31 July 2020 Published 17 August 2020

Academic Editor Ali Kashif Bashir

Copyright copy 2020Maleeha Javed et alis is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Tone mapping operators are designed to display high dynamic range (HDR) images on low dynamic range devices Clustering-based content and color adaptive tone mapping algorithm aims to maintain the color information and local texture However finedetails can still be lost in low dynamic range images is paper presents an effective way of clustering-based content and coloradaptive tone mapping algorithm by using fast search and find of density peak clustering e suggested clustering methodreduces the loss of local structure and allows better adaption of color in images e experiments are carried out to evaluate theeffectiveness and performance of proposed technique with state-of-the-art clustering techniques e objective and subjectiveevaluation results reveal that fast search and find of density peak preserves more textural informationerefore it is most suitableto be used for clustering-based content and color adaptive tone mapping algorithm

1 Introduction

e demand for high dynamic range (HDR) images israpidly increasing with the advent of sensors displaytechnology and availability of multiple exposure photog-raphy HDR images provide better visual quality as com-pared to their counterpart low dynamic range (LDR) imageson account of their large ranged brightness levels and betterpreservation of color variations is enables to store andvisualize bright and dark objects in the same image whichwould normally require multiple images at different expo-sures to be in low dynamic range Visualization of HDR

images on standard display devices (made for displayingLDR images) is challenging as they are incapable of dis-playing the high dynamic range of the image To address thisissue tone mapping operators are used to convert HDRimages to LDR images However the technology that di-rectly displays HDR images is becoming more cost-effectiveRecently most of the available display devices have LDRwhich is why HDR to LDR conversion still has practicalapplications

For visualization of HDR images on LDR devicesRetinex theory has been widely used which slices the imageinto its components base layer and detail layer Edge

HindawiMobile Information SystemsVolume 2020 Article ID 8846033 10 pageshttpsdoiorg10115520208846033

preservation filtering techniques are used for the approxi-mation of base layer in Retinex image decompositionHowever these methods have resulted in false coloring andhalo artifacts because filtering is unable to capture complexlocal structure of the image To overcome this problem weneed a method that can better adopt the local structure andcolor variation of HDR images in LDR images

Furthermore tone mapping operators are generallydivided into two main categories global tone mappingoperators and local tone mapping operators Global tonemapping operators use the same monotonic curve for dy-namic range compression of the entire image [1 2] Initiallythe focus of the research was on the design of global tonemapping operators Drago et al [3] proposed a tonemapping algorithm that used an adaptive logarithmic basefor luminance compression while maintaining the imagersquosdetails and contrast Kim and Kautz [4] introduced a tonemapping operator based on the hypothesis that the humanvisual system adopts Gaussian distribution Reinhard et al[5] proposed the use of spontaneous dodging and burningfor dynamic range compression e terms ldquododgingrdquo andldquoburningrdquo originate from printing where holding backlightfrom a region of print is called dodging and the addition oflight to image is referred to as burning is can be used toincrease or decrease the brightness of a captured imageeffectively Reinhard and Devlin [6] introduced the lightadaptive tone mapping which provides satisfying visualeffects Khan et al [7] introduced a lookup table-basedapproach where they have utilized the histogram of lumi-nance for tone mapping To produce the low dynamic rangeimage appropriate for various viewing conditions Han et al[8] modify Khan et alrsquos [7] method by considering theimpact of ambient light on HVS Generally global tonemapping operators are computationally efficient [9]However they are unable to preserve the local characteristicsand dynamic range variations as they ignore local pixelintensity variations in an image [9 10]

Local tone mapping operators solve the issue of globaltone mapping operators by incorporating the ratio betweenneighboring pixels and compressing each pixel Gu et al [11]and Min et al [12] proposed local tone mapping operatorsbased on layer decomposition where the input image isdivided into base layer and detail layer e layer-decom-position-based approach has also been utilized by authors in[13ndash16] Durand and Dorsey [17] proposed a method thatpreserves the detail layer by encoding large-scale variationsof the base layer Debevec and Gibson [18] proposed amethod that helped to preserve brightness and details ofimage by applying a local luminance adaption function forcompression of dynamic range followed by reinjectingdetails in the low dynamic range image Image color ap-pearance model (iCAM06) [19] employed tone mapping byusing iCAM06 color appearance model by considering theviewing conditions to generate optimal results HoweveriCAM06 [19] led to the poor visual quality due to intro-duction of halo artifact color saturation and gradient re-versal at the edges To eliminate these problems the authorsin [20] proposed an iCAM06 based model In their proposedalgorithm guided filter [20] color adaptive transformation

matrix and HPE primitives were altered to enhance theeffectiveness of model

Krawczyk et al [21] introduced an algorithm based onanchoring theory that decomposed image luminance intopatches and then calculated the lightness for each patch Liet al [22] proposed a symmetrical analysis-synthesis filterfor reducing the intensity range Jia and Zhang [23] used theguided image filter for tone mapping Meylan et al [24]suggested a method based on the characteristic of humanretina to reduce dynamic range while increasing the localcontrast of the image Another retina inspired range com-pression algorithm is used [25] to overcome the halo arti-facts loss of details different visualization across differentdisplays and color saturation

Parraga and Otazu [26] developed a tone mappingmethod based on human color perception by dividing theimage intensity into multiresolution contrast Also they hadused a nonlinear saturation model for dynamic range re-duction based on visual cortex behavior Chen et al [27]introduced Earth moverrsquos distance to segment HDR imageand then applied local tone mapping on each component ofthe image is maintained local texture and balancedperceptional impression To remove artifacts and contrastenhancement Liang et al [28] used nonlinear diffusion l2-based retime model [29] was used for contrast enhancementAhn et al [30] proposed a Retinex-based adaptive local tonemapping algorithm according to which the use of guidedfilter reduced halo artifacts Recently Liang et al [31]proposed a hybrid l1 minus l0 norm-based layer decompositionmodel l1 sparsity term was imposed on the base layer and l0on the detail layer In [32] the authors proposed the use ofdecomposed multiscale Retinex for information preserva-tion in tone-mapped image Shu and Wu [33] employed anoptimal local tone mapping operator to avoid halo artifactsand double edges Rana et al [34] proposed a pixel-wiseadaptive tone mapping operator based on support vectorregression to overcome the problem of drastic illuminationvariation

El Mezeni and Saranovac [35] introduced an enhancedlocal tone mapping operator (ELTM) ELTM decomposedan image into detail and base layers where the base layer wascompressed into both logarithmic and linear domain Localtone mapping operator presented in [36] was operationallysimilar to HVS Li and Zheng [37] presented a saliency-aware local tone mapping operator that preserved edgesusing guided filters Ferradans et al [38] proposed a two-stage algorithm incorporating both global tone mapping andlocal tone mapping In the first step they applied a globaltone mapping operator based on the human visual systemand followed by local method for contrast enhancement inthe second stage To make the local and global tone mappingalgorithm more effective Ambalathankandy et al [39] usedhistogram equalization method implemented on FPGA withefficient resource usage To resolve the overenhancement anonuniform quantization technique is proposed for CTimage enhancement [40] Recently deep convolution neuralnetwork is also used for range compression [41]e authorsin [41] used the output of existing tone mapping operators astraining set and therefore inherited the best properties of all

2 Mobile Information Systems

tone mapping operators Hence they performed robustlywell for all testing images compared to current methods oftone mapping

Li et al [42] introduced a clustering-based content andcolor adaptive tone mapping algorithm that preserved thelocal structure and naturalness of imageey used K-meansalgorithm for the clustering of color structure However K-means is difficult to adopt for a specific problem due toprespecified number of clusters In addition the utilizationof K-means for tone mapping has led to the loss of complexlocal detail in LDR images Furthermore the number ofavailable clustering algorithms encourages the use of themost effective algorithm for clustering-based tone mappingalgorithm erefore to preserve the complex local detailswe have adopted a better clustering technique

In this paper we have proposed a clustering-based tonemapping algorithm to overcome the problems of falsecoloring halo artifacts and loss of complex local structureFor that purpose we have utilized the fast search and find ofdensity peak for content and color adaptive tone mappingis clustering technique automatically recognizes clustersregardless of dimensionality of data e following studycompares various available state-of-the-art clustering algo-rithms for tone mapping operation e effectiveness ofproposed technique is compared with different clusteringalgorithms through subjective and objective evaluationtechniques e experimental analysis suggests that the fastsearch and find of density peak provides visually appealingresult without compromising the local structure of an image

e rest of the paper is composed of the followingsections e second section provides a brief introduction tothe clustering-based content and color adaptive tonemapping algorithm e third section describes differentapplied clustering techniques Furthermore the fourthsection discusses the experimentation conducted to evaluatethe performance of different clustering techniques Ouroverall work is concluded in the fifth section

2 Clustering-BasedContentandColorAdaptiveTone Mapping Algorithm

Li et al [42] proposed clustering-based content and coloradaptive tone mapping algorithm Most of conventionaltone mapping operators split an image into chrominanceand luminance channels Instead Li et al divided an imageinto overlapped color patchese overall algorithm consistsof training and testing phases e algorithm [42] utilizestraining phase to learn PCA transform matrix for each colorstructure cluster from HDR training images while thetesting phase is used to project the HDR test images and tofind the closest match in the training set In the first steplogarithmic transform is applied on each HDR image toenhance the contrast and brightness of low luminance valuesin a channel while compressing the higher ones Afterwardeach image is divided into overlapped color patches to avoidartifacts such as local graying out hue shift or color fringes[43] Each patch is further divided into approximately un-correlated components color structure color variation andpatch mean Patches with varied intensity level may have

similar structures [42] erefore patches are grouped to-gether into different clusters based on color structure andthen PCA transform matrix is obtained for each cluster Inthe testing phase for each patch color structure similaritymeasure is calculated with cluster centers extracted from thetraining data and then relevant PCA matrix is retrieved Fortone mapping S curve arctan function is applied on PCAprojection matrix e patch mean is compressed by a linearfunction and the color variation is controlled by an arctanfunction e image is reconstructed by processed patchesAt the end postprocessing step is performed for contrastenhancement by clamping image at its maximum andminimum intensity value

3 Compared Clustering Techniques

To explore the influence of different clustering algorithmsK-means is a reference clustering algorithm and the resultsare compared with Gaussian Mixture Model (GMM) [44]DBSCAN [45] and fast search and find of density peak(FSFDP) [46] clustering algorithms Before presenting theresults obtained by using each of these clustering schemes abrief description of each clustering technique is presentedwith its strengths and shortcomings

31 K-Means K-means is an extensively used partition-based clustering technique proposed by MacQueen Parti-tioning can be done based on minimizing the objectivefunction known as square error [47] which can be calculatedas follows

J(V) 1113944c

i11113944

ci

j1xi minus vi

1113872 1113873

2 (1)

where xi minus vi ldquocrdquo and ldquocirdquo represented Euclidean distancenumber of cluster centers and the number of points in ith

cluster respectivelyK-means was considered by Li et al [42]for the clustering of the color structure K-means is a rel-atively simple and computationally efficient algorithmHowever it suffers from the restriction of requiring humaninteraction for selecting the number of clusters whichprovides easy implementation and effective results Resultsare influenced by the initialization and may affect theperformance of the tone mapping algorithm FurthermoreK-means is sensitive to the cluster with a single data pointdue to its square distance [48]

32 Fast Search and Find of Density Peak Fast search andfind of density peak (FSFDP) [46] assumes that the density ofdata at the center of a cluster is relatively high and the centerof one cluster is far away from the center of another clusterBased on this assumption FSFDP calculates the number ofclusters and their centers automatically Moreover it detectsand removes outliers intuitively Clusters of nonsphericalshape are easy to be recognized using FSFDP In FSFDP thecutoff distance used for calculating the density of each datapoint has a great influence on the effectiveness of FSFDP For

Mobile Information Systems 3

a data point i FSFDP calculates local density ρi and dis-tance δi from its nearest center as follows

ρi 1113944j

X dij minus dc1113872 1113873 (2)

where dij denotes the distance from data point i to j and dc iscutoff distance used to calculate the density of each datapoint For assigning a data point i to the nearest clustercenter distance is calculated as

δi

minj ρjgtρi

dij1113872 1113873 if existj s tρj gt ρi

maxj ρjgtρi

dij1113872 1113873 otherwise

⎧⎪⎪⎪⎨

⎪⎪⎪⎩

(3)

Equations (2) and (3) represent that the cluster centersare points with greater distance ldquoδrdquo from other clustercenters and high data density ldquoρrdquo e fast search and find ofdensity peak is further experimented on two scale imple-mentations [42] and represented as two-scale fast search andfind of density peak (TFSFDP)

33 Gaussian Mixture Model Unlike K-means clusteringGaussian Mixture Model (GMM) [44] allows clusters ofdifferent shapes e cluster orientation is dependent on theGaussian distribution GMM sets the k number of Gaussiansaccording to the data distribution GMM learns the clusterbelonging through the probability of each data point usingits parameters such as variance Σk mean μk and weight ofthe cluster πk Mathematically it is computed by using

p(x) 1113944K

k1πkX x | μk Σk( 1113857 (4)

GMM allows cluster overlapping however it is com-putationally extensive erefore GMM is not applicable forhigh-dimensional data Moreover GMM is dependent onGaussian distribution (clusters)

34 Density-Based Spatial Clustering of Applications withNoise Similar to FSFDP Density-Based Spatial Clusteringof Application with Noise (DBSCAN) [45] is a density-basedclustering algorithm that does not require the number ofclusters as an initial parameter DBSCAN identifies theclusters of various shapes and sizes based on data connec-tivity Furthermore DBSCAN is widely adopted because ofits efficient computation However the algorithm ofDBSCAN depends on the initial radius ldquorrdquo of the cluster asfollows

Nr i isin D|dist(i j)le r1113864 1113865 (5)

If the value of the radius is set to be small all points thatbelong to the sparse cluster are considered as noise On theother hand if the radius is set to be large all points lie withina single cluster erefore to obtain appropriate results thealgorithm is executed for a number of times with differentvalues of ldquorrdquo

4 Experiments and Results

is section describes the details of the dataset and theevaluation criteria used in our experiments As differentclustering methods have different parameters the details ofthese are discussed in parameter setting section e focus isto elaborate the influence of different state-of-the-art clus-tering algorithms on clustering-based content and coloradaptive tone mapping algorithm which is examined inperformance and effectiveness section

41 Dataset and Evaluation Criteria For the training phaseKodakrsquos database [49] is used which consists of a total of 24images of size of 512times 768times 3 All images are of true color(24 bits per pixel) and for tone mapping these images areused as standard set suit [49] In the testing phase weperformed a series of experiments on Fun and Shi [50] HDRdataset to evaluate the influence of different clustering al-gorithms on clustering-based content and color adaptivetone mapping [42] is dataset contains both indoor andoutdoor HDR images of 105 scenes captured through NikonD700 digital camera Sample images from the dataset areshown in Figure 1

We compared the results of different clustering algo-rithms both qualitatively and quantitatively Tone mappingoperators attempt to preserve desirable characteristics in-cluding local structure and naturalness and to avoid the haloartifacts while converting an HDR image into LDR imagese classical evaluation parameters for objective measure-ments like PSNR cannot be used for the evaluation of tonemapping due to the unavailability of reference LDR imagesSo we used two metrics for quantitative evaluation of dy-namic range reduction with different clustering techniquestone-mapped image quality index (TMQI) [51] and featuresimilarity index for tone-mapped images (FSITM) [52]TMQI [51] is used to compute the structural resemblanceand indexes of naturalness and FSITM [52] is to measurethe local phase similarity between HDR and LDR images

42 Parameters Setting Appearance regeneration of animage is dependent not only on the intensity relationship butalso on many local weighted attributes including brightnesscontrast gray level and color relationship Clustering-basedcontent and color adaptive tone mapping [42] is a patch basemodel in which the size and shape of the patch depend uponthe window used for the calculation of mean and colorstructure In this study we have used the default windowwith size of 7times 7 as defined by Li et al [42] We have useddefault values of the control parameter for color appearancelocal structure and luminance within the tone mappingprocess as used by [42]

For clustering algorithms the control parameters are setwith the most optimal valuese details of these parametersfor each clustering method are as follows

421 For K-means For K-means clustering the number ofclusters is initially set to be 100


422 For Fast Search and Find of Density Peak For fastsearch and find of density peak cut-off distance ldquodcrdquo given inequation (2) is significant to be set However varying itsdefault value that is ldquo14141rdquo has no difference in objectiveevaluation as shown in Table 1 therefore we set it as 14141

423 For Density-Based Spatial Clustering of Applicationswith Noise Furthermore for DBSCAN we execute the

algorithm several times to get to the appropriate value of ldquorrdquoas mentioned in equation (5) On the default value of ldquorrdquowhich is 10 for DBSCAN the algorithm considers sparsedatapoints as noise and therefore selected value of ldquorrdquo to be02

424 For Gaussian Mixture Model For GMM we used thedefault value of Gaussian distribution ldquokrdquo in equation (4) as (2)

Figure 1 HDR tone-mapped source images [50]

148432 154218

471176

637546

0775 0775 0763 07760715 0712 0701 0691

k = 2 k = 8 k = 16 k = 20

Time (sec)FSITMTMQI

Figure 2 Effect of Gaussian distribution ldquokrdquo on time and FSITM and TMQI parameters

Table 1 Effect of cutoff distance on FSITM and TMQI parameters for an image of size 1419times 2130times 3

Objective evaluation dc 13838 dc 14141 dc 19192 dc 23465 dc 28979 dc 32348 dc 4

FSITM 0871 0871 0871 0871 0871 0871 0871TMQI 0881 0881 0881 0881 0881 0881 0881


e experimentation shows that the different values ofGaussian distribution have none or little effect on objectiveevaluation However Gaussian distributionrsquos larger valuescause high computational cost is is illustrated in Figure 2

43 Performance and Effectiveness

431 Objective Evaluation Tables 2 and 3 present the re-sults of different clustering techniques applied with the

method of [42] in terms of TMQI and FSITM respectivelye best results are highlighted in both tables Green blueand yellow numbers represent the first second and thirdbest scores respectively likewise red number represents theworst score e effectiveness of tone mapping operatorsdepends upon the contrast brightness and local structure ofHDR image scenes So a tone mapping operator cannot beequally effective for all HDR images [53] Table 2 depicts thatTFSFDP performs best for 9 images however considering

Table 2 Tone-mapped image quality index score

S no With FSFDP With TFSFDP With K-means With GMM With DBSCAN1 0758 0761 0766 0715 03732 0772 0780 0774 0723 02293 0818 0835 0821 0646 04904 0824 0820 0829 0781 04905 0823 0830 0826 0790 03656 0881 0893 0878 0820 04607 0762 0781 0754 0778 02588 0727 0730 0728 0625 04019 0880 0889 0878 0800 048610 0789 0804 0783 0744 042611 0917 0916 0907 0769 052012 0874 0873 0880 0818 0358Avg 0819 0826 0819 0683 0405

Table 3 Feature similarity index for tone-mapped image score


0 500 1000 1500

With FSFDP

With K-means

With GMM

WithDBSCAN

Time (sec)

Figure 3 Average execution time with different clustering algorithm applied on five images with size of 535times 401times 3


the average of each method K-means and FSFDP can beused alternatively Table 3 has strengthened the resultsfurther as FSITM shows that TFSFDP performs best whileK-means and FSFDP stood in the second and third posi-tions respectively Furthermore DBSCAN performs worstin terms of both FSITM and TMQI for every image Basedon the obtained results DBSCAN is not recommended fortone mapping

Performance of clustering-based tone mapping techniquesis also compared in terms of their execution time e exe-cution times of different clustering-based tone mappingtechniques are presented in Figure 3 which highlights thatGMM has high computation cost irrespective of the providedTMQI and FSITM results erefore it is not a preferredclustering algorithm to be used for tone mapping operation Interms of fastest execution timeK-means outperforms the other

(a) (b) (c)

(d) (e)

Figure 4 Visual comparison of outdoor tone-mapped images with (a) FSFDP (b) TFSFDP (c) K-means (d) GMM and (e) DBSCAN

(a) (b) (c)

(d) (e)

Figure 5 Visual comparison of indoor tone-mapped images with (a) FSFDP (b) TFSFDP (c) K-means (d) GMM and (e) DBSCAN


methods with least execution time of about ldquo6598 secondsrdquowhereas DBSCAN and FSFDP can also be considered for tonemapping operation as their execution times are about ldquo1131secondsrdquo and ldquo12218 secondsrdquo respectively

432 Subjective Evaluation Assessment Subjective evalua-tion is performed through visual assessment and scoringFigures 4 and 5 exhibit a visual comparison of clustering-based techniques using tone-mapped indoor and outdoorscene images FSFDP TFSFDP and K-means produced thevisually appealing results as these clustering methods betterpreserve the local structure and color saturation as shown inFigures 4(a)ndash4(c) and Figures 5(a)ndash5(c) In case of FSFDPminor details like clouds near the shaded corner shown inFigures 4(a) and 4(b) and wood pattern of cutting board as inFigures 5(a) and 5(b) remain better preserved than K-meansthat can be clearly illustrated from Figures 4(c) and 5(c)GMM suffers from information loss and unpleasant effectIn Figures 4(d) and 5(d) details on a glass of the window andshelve are not visible In case of DBSCAN colors areoversaturated and result in unnatural visual appearances ofimage as depicted in Figures 4(e) and 5(e)

Subjective evaluation of 12 images is performed by 8volunteers currently working on image processing andmachine learning We used a monitor with a spatial reso-lution of 2560times1600 to display images side by side using theimage viewer customizable window We showed the re-sultant images of all methods to volunteers and asked themto score image from 1 to 10 ldquo1rdquo represents the worst scorewhile ldquo10rdquo is the highest score Ratings are recorded by usingscore sheets provided to each volunteer during the subjectiveassessment Figure 6 shows that for most of the imagesFSFDP has the highest averaged score

5 Conclusion

Tone mapping is a complex field in which the conversionof HDR image to LDR image without information loss is

needed e intention of this article is to present the bestclustering technique among existing techniques for pre-serving complex local details lost in case of clustering-based content and color adaptive tone mapping methodFor this purpose the influence of different clusteringtechniques is examined e effectiveness is measured interms of subjective evaluation and FSITM and TMQIvalues Experiments show that fast search and find ofdensity peak results in more appealing results for tonemapping with acceptable computational cost In the fu-ture this model can be extended by using a feature otherthan color structure for clustering of patches Besides anyother function in place of S shape arctan curve can be usedfor range compression

Data Availability

e data used to support the findings of this study areavailable from the corresponding author upon request

Conflicts of Interest

e authors declare that there are no conflicts of interestregarding the publication of this paper

References

[1] J Tumblin and H Rushmeier ldquoTone reproduction for realisticimagesrdquo IEEE Computer Graphics and Applications vol 13no 6 pp 42ndash48 1993

[2] G Ward ldquoA contrast-based scalefactor for luminance dis-playrdquo in Graphics Gems vol 4 pp 415ndash421 ElsevierAmsterdam Netherlands 1994

[3] F Drago K Myszkowski T Annen and N Chiba ldquoAdaptivelogarithmic mapping for displaying high contrast scenesrdquoComputer Graphics Forum vol 22 no 3 pp 419ndash426 2003

[4] M H Kim and J Kautz ldquoConsistent tone reproductionrdquo inProceedings of the Proceedings of Computer Graphics andImaging Innsbruck Austria February 2008

0

1

2

3

4

5

6

7

8

9

10

1 2 3 4 5 6 7 8 9 10 11 12

With FSFDBWith TFSFDPWith K-means

With GMMWith DbScan

Figure 6 Average subjective evaluation score


[5] E Reinhard M Stark P Shirley and J Ferwerda ldquoPhoto-graphic tone reproduction for digital imagesrdquo ACM Trans-actions on Graphics (TOG) vol 21 no 3 pp 267ndash276 2002

[6] E Reinhard and K Devlin ldquoDynamic range reduction in-spired by photoreceptor physiologyrdquo IEEE Transactions onVisualization and Computer Graphics vol 11 no 1 pp 13ndash242005

[7] I R Khan S Rahardja M M Khan M M Movania andF Abed ldquoA tone-mapping technique based on histogramusing a sensitivity model of the human visual systemrdquo IEEETransactions on Industrial Electronics vol 65 no 4pp 3469ndash3479 2018

[8] J Han I R Khan and S Rahardja ldquoLighting conditionadaptive tone mapping methodrdquo in Proceedings of the ACMSIGGRAPH 2018 Posters onmdashSIGGRAPHrsquo18 pp 1-2 Van-couver Canada August 2018

[9] D Lischinski Z Farbman M Uyttendaele and R SzeliskildquoInteractive local adjustment of tonal valuesrdquo in ACMTransactions on Graphics vol 25 no 3 pp 646ndash653 ACM2006

[10] D-H Lee M Fan S-W Kim M-C Kang and S-J KoldquoHigh dynamic range image tone mapping based on asym-metric model of retinal adaptationrdquo Signal Processing ImageCommunication vol 68 pp 120ndash128 2018

[11] B Gu W Li M Zhu and M Wang ldquoLocal edge-preservingmultiscale decomposition for high dynamic range image tonemappingrdquo IEEE Transactions on Image Processing vol 22no 1 pp 70ndash79 2013

[12] D Min S Choi J Lu B Ham K Sohn and M N Do ldquoFastglobal image smoothing based on weighted least squaresrdquoIEEE Transactions on Image Processing vol 23 no 12pp 5638ndash5653 2014

[13] Z Farbman R Fattal D Lischinski and R Szeliski ldquoEdge-preserving decompositions for multi-scale tone and detailmanipulationrdquo ACM Transactions on Graphics (TOG) ACMvol 27 no 3 p 67 2008

[14] S Paris S W Hasinoff and J Kautz ldquoLocal Laplacian filtersrdquoCommunications of the ACM vol 58 no 3 pp 81ndash91 2015

[15] L Xu C Lu Y Xu and J Jia ldquoImage smoothing via L 0gradient minimizationrdquo in ACM Transactions on Graphics(TOG) vol 30 no 6 p 174 ACM 2011

[16] L Xu Q Yan Y Xia and J Jia ldquoStructure extraction fromtexture via relative total variationrdquo ACM Transactions onGraphics (TOG) vol 31 no 6 p 139 2012

[17] F Durand and J Dorsey ldquoFast bilateral filtering for the displayof high-dynamic-range imagesrdquo ACM Transactions onGraphics (TOG) ACM vol 21 no 3 pp 257ndash266 2002

[18] P Debevec and S Gibson ldquoA tone mapping algorithm forhigh contrast imagesrdquo in Proceedings of the 13th EurographicsWorkshop on Rendering Pisa Italy June 2002

[19] J Kuang G M Johnson and M D Fairchild ldquoiCAM06 arefined image appearance model for HDR image renderingrdquoJournal of Visual Communication and Image Representationvol 18 no 5 pp 406ndash414 2007

[20] S Tong and Y Yang ldquoA novel tone mapping algorithmrdquo inProceedings of the 2019 IEEE 8th Joint International Infor-mation Technology and Artificial Intelligence Conference(ITAIC) pp 427ndash431 IEEE Chongqing China May 2019

[21] G Krawczyk K Myszkowski and H P Seidel ldquoLightnessperception in tone reproduction for high dynamic rangeimagesrdquo Computer Graphics Forum vol 24 no 3 pp 635ndash645 2005

[22] Y Li L Sharan and E H Adelson ldquoCompressing andcompanding high dynamic range images with subband

architecturesrdquo ACM Transactions on Graphics vol 24 no 3pp 836ndash844 2005

[23] Y Jia and W Zhang ldquoEfficient and adaptive tone mappingalgorithm based on guided image filterrdquo International Journalof Pattern Recognition and Artificial Intelligence vol 34 no 4p 2054012 2020

[24] L Meylan D Alleysson and S Susstrunk ldquoModel of retinallocal adaptation for the tone mapping of color filter arrayimagesrdquo Journal of the Optical Society of America A vol 24no 9 pp 2807ndash2816 2007

[25] X-S Zhang K-F Yang J Zhou and Y-J Li ldquoRetina inspiredtone mappingmethod for high dynamic range imagesrdquoOpticsExpress vol 28 no 5 pp 5953ndash5964 2020

[26] C A Parraga and X Otazu ldquoWhich tone-mapping operator isthe best A comparative study of perceptual qualityrdquo Journalof the Optical Society of America A vol 35 no 4 pp 626ndash6382018

[27] H T Chen T L Liu and C S Fuh ldquoTone reproduction aperspective from luminance-driven perceptual groupingrdquoInternational Journal of Computer Vision vol 65 no 1-2pp 73ndash96 2005

[28] Z Liang W Liu and R Yao ldquoContrast enhancement bynonlinear diffusion filteringrdquo IEEE Transactions on ImageProcessing vol 25 no 2 pp 673ndash686 2016

[29] M K Ng andWWang ldquoA total variation model for RetinexrdquoSIAM Journal on Imaging Sciences vol 4 no 1 pp 345ndash3652011

[30] H Ahn B Keum D Kim and H S Lee ldquoAdaptive local tonemapping based on retinex for high dynamic range imagesrdquo inProceedings of the 2013 IEEE International Conference onConsumer Electronics (ICCE) pp 153ndash156 IEEE Las VegasNV USA January 2013

[31] Z Liang J Xu D Zhang Z Cao and L Zhang ldquoA hybrid l1-l0 layer decomposition model for tone mappingrdquo in Pro-ceedings of the IEEE Conference on Computer Vision andPattern Recognition pp 4758ndash4766 Salt Lake City UT USAJune 2018

[32] B J Lee and B C Song ldquoLocal tone mapping using sub-banddecomposed multi-scale retinex for high dynamic rangeimagesrdquo in Proceedings of the 2014 IEEE International Con-ference on Consumer Electronics (ICCE) pp 125ndash128 LasVegas NV USA January 2014

[33] X Shu and X Wu ldquoLocally adaptive rank-constrained op-timal tone mappingrdquo ACM Transactions on Graphics (TOG)vol 37 no 3 pp 1ndash10 2018

[34] A Rana G Valenzise and F Dufaux ldquoLearning-basedadaptive tone mapping for keypoint detectionrdquo in Proceedingsof the 2017 IEEE International Conference on Multimedia andExpo (ICME) pp 337ndash342 Hong Kong China July 2017

[35] D M El Mezeni and L V Saranovac ldquoEnhanced local tonemapping for detail preserving reproduction of high dynamicrange imagesrdquo Journal of Visual Communication and ImageRepresentation vol 53 pp 122ndash133 2018

[36] S A Henley ldquoRendering of high dynamic range imagesrdquo USPatent No 7480421 US Patent and Trademark OfficWashington DC USA 2009

[37] Z Li and J Zheng ldquoVisual-salience-based tone mapping forhigh dynamic range imagesrdquo IEEE Transactions on IndustrialElectronics vol 61 no 12 pp 7076ndash7082 2014

[38] S Ferradans M Bertalmio E Provenzi and V Caselles ldquoAnanalysis of visual adaptation and contrast perception for tonemappingrdquo IEEE Transactions on Pattern Analysis and Ma-chine Intelligence vol 33 no 10 pp 2002ndash2012 2011


[39] P Ambalathankandy M Ikebe T Yoshida et al ldquoAnadaptive global and local tone mapping algorithm imple-mented on FPGArdquo IEEE Transactions on Circuits and Systemsfor Video Technology 2019

[40] A Mehmood I R Khan D Hassan and D Hussain ldquoEn-hancement of CT images for visualizationrdquo in Proceedings ofthe ACM SIGGRAPH 2019 Posters vol 83 pp 1-2 New YorkNY USA July 2019

[41] A Rana P Singh G Valenzise F Dufaux N Komodakis andA Smolic ldquoDeep tone mapping operator for high dynamicrange imagesrdquo IEEE Transactions on Image Processing vol 29pp 1285ndash1298 2019

[42] H Li X Jia and L Zhang ldquoClustering based content andcolor adaptive tone mappingrdquo Computer Vision and ImageUnderstanding vol 168 pp 37ndash49 2018

[43] D Ellis Tone Mapping for High Dynamic Range CamerasDepartment of Engineering Science Oxford University Ox-ford UK 2008 httpwwwrobotsoxacuksimdredocsellis_d_stj_4yppdf

[44] Clustering with Gaussian mixture modelmdashclustering withGaussian mixture modelmdashmedium 2018 httpsmediumcomclustering-with-gaussian-mixture-modelclustering-with-gaussian-mixture-model-c695b6cd60da

[45] M Ester H P Kriegel J Sander and X Xu ldquoA density-basedalgorithm for discovering clusters in large spatial databaseswith noiserdquo Computer Science vol 96 no 34 pp 226ndash2311996

[46] A Rodriguez and A Laio ldquoClustering by fast search and findof density peaksrdquo Science vol 344 no 6191 pp 1492ndash14962014

[47] S A Elavarasi J Akilandeswari and B Sathiyabhama ldquoAsurvey on partition clustering algorithmsrdquo InternationalJournal of Enterprise Computing and Business Systems vol 1no 1 2011

[48] S K Popat and M Emmanuel ldquoReview and comparativestudy of clustering techniquesrdquo International Journal ofComputer Science and Information Technologies vol 5 no 1pp 805ndash812 2014

[49] httpr0kusgraphicskodak 2018[50] B Funt and L Shi ldquoHDR datasetrdquo 2018 httpwwwcssfuca

simcolourdatafunt_hdr[51] H Yeganeh and Z Wang ldquoObjective quality assessment of

tone-mapped imagesrdquo IEEE Transactions on Image Process-ing vol 22 no 2 pp 657ndash667 2013

[52] H Z Nafchi A Shahkolaei R F Moghaddam andM Cheriet ldquoFSITM a feature similarity index for tone-mapped imagesrdquo IEEE Signal Processing Letters vol 22 no 8pp 1026ndash1029 2015

[53] K Gu S Wang G Zhai et al ldquoBlind quality assessment oftone-mapped images via analysis of information naturalnessand structurerdquo IEEE Transactions on Multimedia vol 18no 3 pp 432ndash443 2016


preservation filtering techniques are used for the approxi-mation of base layer in Retinex image decompositionHowever these methods have resulted in false coloring andhalo artifacts because filtering is unable to capture complexlocal structure of the image To overcome this problem weneed a method that can better adopt the local structure andcolor variation of HDR images in LDR images

Furthermore tone mapping operators are generallydivided into two main categories global tone mappingoperators and local tone mapping operators Global tonemapping operators use the same monotonic curve for dy-namic range compression of the entire image [1 2] Initiallythe focus of the research was on the design of global tonemapping operators Drago et al [3] proposed a tonemapping algorithm that used an adaptive logarithmic basefor luminance compression while maintaining the imagersquosdetails and contrast Kim and Kautz [4] introduced a tonemapping operator based on the hypothesis that the humanvisual system adopts Gaussian distribution Reinhard et al[5] proposed the use of spontaneous dodging and burningfor dynamic range compression e terms ldquododgingrdquo andldquoburningrdquo originate from printing where holding backlightfrom a region of print is called dodging and the addition oflight to image is referred to as burning is can be used toincrease or decrease the brightness of a captured imageeffectively Reinhard and Devlin [6] introduced the lightadaptive tone mapping which provides satisfying visualeffects Khan et al [7] introduced a lookup table-basedapproach where they have utilized the histogram of lumi-nance for tone mapping To produce the low dynamic rangeimage appropriate for various viewing conditions Han et al[8] modify Khan et alrsquos [7] method by considering theimpact of ambient light on HVS Generally global tonemapping operators are computationally efficient [9]However they are unable to preserve the local characteristicsand dynamic range variations as they ignore local pixelintensity variations in an image [9 10]

Local tone mapping operators solve the issue of globaltone mapping operators by incorporating the ratio betweenneighboring pixels and compressing each pixel Gu et al [11]and Min et al [12] proposed local tone mapping operatorsbased on layer decomposition where the input image isdivided into base layer and detail layer e layer-decom-position-based approach has also been utilized by authors in[13ndash16] Durand and Dorsey [17] proposed a method thatpreserves the detail layer by encoding large-scale variationsof the base layer Debevec and Gibson [18] proposed amethod that helped to preserve brightness and details ofimage by applying a local luminance adaption function forcompression of dynamic range followed by reinjectingdetails in the low dynamic range image Image color ap-pearance model (iCAM06) [19] employed tone mapping byusing iCAM06 color appearance model by considering theviewing conditions to generate optimal results HoweveriCAM06 [19] led to the poor visual quality due to intro-duction of halo artifact color saturation and gradient re-versal at the edges To eliminate these problems the authorsin [20] proposed an iCAM06 based model In their proposedalgorithm guided filter [20] color adaptive transformation

matrix and HPE primitives were altered to enhance theeffectiveness of model

Krawczyk et al [21] introduced an algorithm based onanchoring theory that decomposed image luminance intopatches and then calculated the lightness for each patch Liet al [22] proposed a symmetrical analysis-synthesis filterfor reducing the intensity range Jia and Zhang [23] used theguided image filter for tone mapping Meylan et al [24]suggested a method based on the characteristic of humanretina to reduce dynamic range while increasing the localcontrast of the image Another retina inspired range com-pression algorithm is used [25] to overcome the halo arti-facts loss of details different visualization across differentdisplays and color saturation

Parraga and Otazu [26] developed a tone mappingmethod based on human color perception by dividing theimage intensity into multiresolution contrast Also they hadused a nonlinear saturation model for dynamic range re-duction based on visual cortex behavior Chen et al [27]introduced Earth moverrsquos distance to segment HDR imageand then applied local tone mapping on each component ofthe image is maintained local texture and balancedperceptional impression To remove artifacts and contrastenhancement Liang et al [28] used nonlinear diffusion l2-based retime model [29] was used for contrast enhancementAhn et al [30] proposed a Retinex-based adaptive local tonemapping algorithm according to which the use of guidedfilter reduced halo artifacts Recently Liang et al [31]proposed a hybrid l1 minus l0 norm-based layer decompositionmodel l1 sparsity term was imposed on the base layer and l0on the detail layer In [32] the authors proposed the use ofdecomposed multiscale Retinex for information preserva-tion in tone-mapped image Shu and Wu [33] employed anoptimal local tone mapping operator to avoid halo artifactsand double edges Rana et al [34] proposed a pixel-wiseadaptive tone mapping operator based on support vectorregression to overcome the problem of drastic illuminationvariation

El Mezeni and Saranovac [35] introduced an enhancedlocal tone mapping operator (ELTM) ELTM decomposedan image into detail and base layers where the base layer wascompressed into both logarithmic and linear domain Localtone mapping operator presented in [36] was operationallysimilar to HVS Li and Zheng [37] presented a saliency-aware local tone mapping operator that preserved edgesusing guided filters Ferradans et al [38] proposed a two-stage algorithm incorporating both global tone mapping andlocal tone mapping In the first step they applied a globaltone mapping operator based on the human visual systemand followed by local method for contrast enhancement inthe second stage To make the local and global tone mappingalgorithm more effective Ambalathankandy et al [39] usedhistogram equalization method implemented on FPGA withefficient resource usage To resolve the overenhancement anonuniform quantization technique is proposed for CTimage enhancement [40] Recently deep convolution neuralnetwork is also used for range compression [41]e authorsin [41] used the output of existing tone mapping operators astraining set and therefore inherited the best properties of all












J(V) 1113944c

i11113944

ci

j1xi minus vi

1113872 1113873

2 (1)






ρi 1113944j

X dij minus dc1113872 1113873 (2)


δi

minj ρjgtρi


maxj ρjgtρi


⎧⎪⎪⎪⎨

⎪⎪⎪⎩

(3)



p(x) 1113944K

k1πkX x | μk Σk( 1113857 (4)


















148432 154218

471176

637546

0775 0775 0763 07760715 0712 0701 0691

k = 2 k = 8 k = 16 k = 20

Time (sec)FSITMTMQI




FSITM 0871 0871 0871 0871 0871 0871 0871TMQI 0881 0881 0881 0881 0881 0881 0881










0 500 1000 1500

With FSFDP

With K-means

With GMM

WithDBSCAN

Time (sec)





(a) (b) (c)

(d) (e)


(a) (b) (c)

(d) (e)






5 Conclusion



Data Availability




References





0

1

2

3

4

5

6

7

8

9

10

1 2 3 4 5 6 7 8 9 10 11 12


With GMMWith DbScan

































































J(V) 1113944c

i11113944

ci

j1xi minus vi

1113872 1113873

2 (1)






ρi 1113944j

X dij minus dc1113872 1113873 (2)


δi

minj ρjgtρi


maxj ρjgtρi


⎧⎪⎪⎪⎨

⎪⎪⎪⎩

(3)



p(x) 1113944K

k1πkX x | μk Σk( 1113857 (4)


















148432 154218

471176

637546

0775 0775 0763 07760715 0712 0701 0691

k = 2 k = 8 k = 16 k = 20

Time (sec)FSITMTMQI




FSITM 0871 0871 0871 0871 0871 0871 0871TMQI 0881 0881 0881 0881 0881 0881 0881










0 500 1000 1500

With FSFDP

With K-means

With GMM

WithDBSCAN

Time (sec)





(a) (b) (c)

(d) (e)


(a) (b) (c)

(d) (e)






5 Conclusion



Data Availability




References





0

1

2

3

4

5

6

7

8

9

10

1 2 3 4 5 6 7 8 9 10 11 12


With GMMWith DbScan
























































ρi 1113944j

X dij minus dc1113872 1113873 (2)


δi

minj ρjgtρi


maxj ρjgtρi


⎧⎪⎪⎪⎨

⎪⎪⎪⎩

(3)



p(x) 1113944K

k1πkX x | μk Σk( 1113857 (4)


















148432 154218

471176

637546

0775 0775 0763 07760715 0712 0701 0691

k = 2 k = 8 k = 16 k = 20

Time (sec)FSITMTMQI




FSITM 0871 0871 0871 0871 0871 0871 0871TMQI 0881 0881 0881 0881 0881 0881 0881










0 500 1000 1500

With FSFDP

With K-means

With GMM

WithDBSCAN

Time (sec)





(a) (b) (c)

(d) (e)


(a) (b) (c)

(d) (e)






5 Conclusion



Data Availability




References





0

1

2

3

4

5

6

7

8

9

10

1 2 3 4 5 6 7 8 9 10 11 12


With GMMWith DbScan




























































148432 154218

471176

637546

0775 0775 0763 07760715 0712 0701 0691

k = 2 k = 8 k = 16 k = 20

Time (sec)FSITMTMQI




FSITM 0871 0871 0871 0871 0871 0871 0871TMQI 0881 0881 0881 0881 0881 0881 0881










0 500 1000 1500

With FSFDP

With K-means

With GMM

WithDBSCAN

Time (sec)





(a) (b) (c)

(d) (e)


(a) (b) (c)

(d) (e)






5 Conclusion



Data Availability




References





0

1

2

3

4

5

6

7

8

9

10

1 2 3 4 5 6 7 8 9 10 11 12


With GMMWith DbScan































































0 500 1000 1500

With FSFDP

With K-means

With GMM

WithDBSCAN

Time (sec)





(a) (b) (c)

(d) (e)


(a) (b) (c)

(d) (e)






5 Conclusion



Data Availability




References





0

1

2

3

4

5

6

7

8

9

10

1 2 3 4 5 6 7 8 9 10 11 12


With GMMWith DbScan

























































(a) (b) (c)

(d) (e)


(a) (b) (c)

(d) (e)






5 Conclusion



Data Availability




References





0

1

2

3

4

5

6

7

8

9

10

1 2 3 4 5 6 7 8 9 10 11 12


With GMMWith DbScan


























































5 Conclusion



Data Availability




References





0

1

2

3

4

5

6

7

8

9

10

1 2 3 4 5 6 7 8 9 10 11 12


With GMMWith DbScan























































personalcommunicationtechnologiesforsmartspaces...

Documents