Dodo6/Topic_Modelling_using_LDA
1
1electronics2Article3 4Multiscale Image Matting Based Multi-Focus Image5Fusion Technique6Sarmad Maqsood 1,2 , Umer Javed 1 , Muhammad Mohsin Riaz 3 , Muhammad Muzammil 1 ,7Fazal Muhammad 2 and Sunghwan Kim 4, *819210311412 13*14 15Faculty of Engineering and Technology, International Islamic University, Islamabad 44000, Pakistan;16sarmad.maqsood@cusit.edu.pk (S.M.); umer.javed@iiu.edu.pk (U.J.); m.muzammil@iiu.edu.pk (M.M.)17Department of Electrical Engineering, City University of Science and Information Technology,18Peshawar 25000, Pakistan; fazal.muhammad@cusit.edu.pk19Center for Advanced Studies in Telecommunication, COMSATS University, Islamabad 44000, Pakistan;20mohsin.riaz@comsats.edu.pk21School of Electrical Engineering, University of Ulsan, Ulsan 44610, Korea22Correspondence: sungkim@ulsan.ac.kr23 24Received: 2 February 2020; Accepted: 8 March 2020; Published: 12 March 202025 262728 29Abstract: Multi-focus image fusion is a very essential method of obtaining an all focus image from30multiple source images. The fused image eliminates the out of focus regions, and the resultant image31contains sharp and focused regions. A novel multiscale image fusion system based on contrast32enhancement, spatial gradient information and multiscale image matting is proposed to extract the33focused region information from multiple source images. In the proposed image fusion approach,34the multi-focus source images are firstly refined over an image enhancement algorithm so that the35intensity distribution is enhanced for superior visualization. The edge detection method based36on a spatial gradient is employed for obtaining the edge information from the contrast stretched37images. This improved edge information is further utilized by a multiscale window technique to38produce local and global activity maps. Furthermore, a trimap and decision maps are obtained39based upon the information provided by these near and far focus activity maps. Finally, the fused40image is achieved by using an enhanced decision maps and fusion rule. The proposed multiscale41image matting (MSIM) makes full use of the spatial consistency and the correlation among source42images and, therefore, obtains superior performance at object boundaries compared to region-based43methods. The achievement of the proposed method is compared with some of the latest techniques44by performing qualitative and quantitative evaluation.45Keywords: image fusion; multi-focus; trimaps; focus maps46 471. Introduction48During image acquisition, one of the most important objectives is obtaining a focused region49of interest. However, because of the limited field depth, the focused region contains sharp edges,50whereas the other regions get blurred. Recently, multi-focus image fusion (combine images with51different focused objects) has received tremendous attention amongst the researchers. This fused52image offers high quality containing more detailed information [1,2]. Several methods are developed53to fuse multiple images, which are broadly grouped into transform and spatial domains [3,4].54Transform domain methods fuse the corresponding transform coefficients and employ inverse55transformation to construct the fused image. Spatial domain methods are further classified into56pixel [5,6] and region based methods [7,8]. The spatial domain methods form the fuse image by57choosing the pixels/regions/blocks that are focused. Transform domain-based methods in dynamic58 59Electronics 2020, 9, 472; doi:10.3390/electronics903047260 61www.mdpi.com/journal/electronics62 63 Electronics 2020, 9, 47264 652 of 1666 67scenes merge these coefficients without considering the spatial properties, resulting in artifacts in the68fused image. Furthermore, pixel and region-based methods are unable to produce the best fusion69results for images with complicated texture patterns [1].70Zhang et al. [9] used morphological operations to extract focus regions. However, this technique71suffers from block artifacts. De et al. [10] utilized morphological processes to detect the focused region72and suggested a technique for calculating an optimized block size. The fused result still suffers from73blocking effects. Later on, Bai et al. [11] presented a novel quadtree decomposition and a weighted74focus based image fusion technique. However, this technique also provides inaccurate segmentation75and low visual effects because of the smooth regions. Yin et al. [12] proposed a method based on76joint dictionary and singular value decomposition (SVD) methods. Still, this method is not effective77computationally because of the individual training for sub dictionaries and SVD computation.78Li et al. [13] explored guided filtering (GFF) and spatial information to improve the fusion results79by mitigating the block effects. Zhang et al. [14] proposed a multifocus image scheme based upon80a visual saliency method. Recently, image matting has been used for effectively differentiating the81focused and out-of-focus regions. These methods can be broadly categorized as supervised matting82and unsupervised matting techniques. Supervised methods require a user specified foreground and83background regions known as trimap. Therefore, such techniques require human experts, are time84consuming, and produce inconsistent results for images with high-textured backgrounds. However,85unsupervised methods are better than supervised ones because user interaction is not required for86achieving a good matting result. Chen et al. [15] used a parametric edge based method. However,87these methods do not consider the artifacts among the smooth regions, and much of it depends88on the performance of hand crafted features, which require much expert knowledge. Li et al. [16]89proposed multifocus matting (MFM) based image fusion by combining together the focus region and90its neighboring pixels. This method marginally improves the fusion results and also overcomes some91shortcomings of spatial domain methods.92Xiao et al. [17] used depth information to segment an image into focus and blur regions.93Zhang et al. [18] made use of log spectrum, Fourier transform, and Bayesian techniques. In [19],94a definite focus region is detected by using a novel multi scale gradient information. Liu et al. [20]95proposed a transform (which is scale invariant) to detect focus regions. However, this technique fails96to offer sharp edges of the focus regions. Furthermore, in [21], the focus information was extracted by97using texture features. Baohua et al. [22] performed the near and far focus region detection by using a98sparse representation and guided filter techniques. In [23], a structure tensor was used for the detection99of high and low frequency components. However, this technique fails to provide a visible difference100between focus and defocus regions in many cases. Yu at.al. [24] presented a convolutional neural101network (CNN) based multifocus image fusion technique. However, in this method, the precision of102recognizing the focus block is very low.103In this paper, a novel multi-focus image fusion method is presented using contrast stretching104and spatial gradients to enhance the edges from the source images. A multiscale sliding window105method is used for detecting the local and global intensity variations to generate initial activity maps.106These multiple activity maps are further processed to generate a trimap. An enhanced image matting107technique is used for generating the decision maps. Finally, the fused image is obtained after processing108the source images, enhanced decision maps, and employing the fusion rule.1092. Proposed Fusion Technique110The schematic diagram of the proposed algorithm is shown in Figure 1. It can be observed that111in the first step, a contrast enhancement scheme is applied on the source images. In the second step,112the outcome of the intensity transformed image is processed through an edge detection method. In a113multi-focus image fusion scheme, the selection of near focus and far focus region plays a vital role.114The region that is in focus during image acquisition tends to have sharp edges as compared to the115 116 Electronics 2020, 9, 472117 1183 of 16119 120out-of-focus region. Therefore, these sharp edges can be detected easily by applying an appropriate121edge detection method.122The edge detection schemes rely heavily on the intensity distribution of an image. A poor123intensity distribution can lead to an oversaturated, undersaturated, dark, or bright image. In either124of those images, the edge detection algorithm cannot perform well. In order to improve the intensity125distribution of an image, an intensity transformation can be performed. In Figure 2, the improvement126in edge information is shown by comparing the images before and after applying the contrast127enhancement scheme. In the next step, a sliding window technique with two different scales is128applied on both edges of the detected images to generate activity maps. In this step, both local and129global intensity variations are analyzed. The fine details are more prominent under a small sliding130window scale. These masks are further fused together and processed to generate a trimap. Next,131the trimap undergoes an image matting transformation to produce refined decision maps, which132produces the final fused image. The proposed fusion scheme, along with the equation references,133is also elaborated in Algorithm 1.134Let Ii be the source color images with M × N dimensions where, m = 1, 2, . . . , M, n = 1, 2, . . . , N135and i ∈ [1, 2] represents near and far focus images, respectively.136 137Figure 1. Schematic diagram of the proposed approach for the image fusion algorithm.138 139Algorithm 1: Proposed MSIM based Fusion Technique.140Require: Ii , i ∈ [1, 2].141Step 1. Apply contrast enhancement on Ii using Equation (1).142Step 2. Apply edge map on Íi using Equation (2) to Equation (5).143σ=1144Step 3. Compute activity maps, G´i,σ ←−−− w, Zi using Equation (6).145Sum Filter, σ = 1146 147Step 4. Compute Smooth Activity maps, Gi,σ ←−−−−−−−−− G´i,σ using Equation (7).148σ=1149 150Step 5. Compute Score maps, ζ i,σ ←−−− Gi,σ using Equations (8) and (9).151Step 6. Repeat Steps 3, 4, and 5 with σ = 3.152Step 7. Compute Near focus (D1 ) and Far focus (D2 ) using Equation (10).153Step 8. Generate Trimap T using Equation (11).154Step 9. Generate Alpha Matte α using Equation (14).155Step 10. Generate Fused Image A F using Equation (15).156 157 Electronics 2020, 9, 472158 1594 of 16160 1612.1. Contrast Enhancement162Improving the enhancement of the low contrast image, the histogram equalization seems an163effective method. Non-parametric modified histogram equalization (NMHE) [25] is integrated to164enhance the contrast and preserves the mean brightness of the source image Ii , i.e.,165NMHE166I´i ←−−−−−− Ii167 168(1)169 170Image development in contrast centrally improves and concentrates pixel details. Figure 2171shows the enhancement in edge information. Figure 2a,b shows the far and near focus source172image, respectively, and their gradients are shown in Figure 2c,d. Contrast enhanced of near and far173focus images are displayed in Figure 2e,f, and their respective edge maps are shown in Figure 2g,h,174respectively. From the images, it can be clearly seen that after the enhancement algorithm, the gradients175of the source image were greatly improved.176 177(a)178 179(b)180 181(c)182 183(d)184 185(e)186 187(f)188 189(g)190 191(h)192 193Figure 2. Results of edge detection after contrast enhancement. (a) Near focus image, (b) far focus194image, (c,d) gradients of (a,b) achieved by a spatial stimuli sketch model (SSGSM) [26], (e,f) contrast195enhancement using non-parametric modified histogram equalization (NMHE) [25], (g,h) gradients of196(e,f) achieved by SSGSM [26].197 1982.2. Edge Detection199The edges of the images after contrast enhancement is done by a spatial stimuli sketch model200(SSGSM) [26] technique, which principally focuses on focal intensity points and edges in an image,201and then the unknown region is calculated in the coarse decision maps by implementing the202concentrated information in both the activity level maps. The weight of the local stimuli is deliberated203by detecting the local variation in the perceived brightness at the respective positions. The discerned204brightness, Pi of a specific image is given in Equation (2) as,205Pi = ϑlog10 ( I´i )206 207(2)208 209where, I´i represents the source images, and ϑ denotes the scaling factor.210Gradients illustrate the sharp intensity variations in the image. Mathematically, the weight is211computed as the total difference of the perceived brightness on x and y directions. The intensity212y213variations of Pi on the x and y axis are represented by $ix and $i , respectively. These variations are214y215calculated by using their respective gradients Bix and Bi , given as in Equations (3) and (4):216y217 218gradient219 220[ Bix , Bi ] ←−−−−−−− Pi221 222(3)223 224 Electronics 2020, 9, 472225 2265 of 16227 228x229 230y231 232y233 234y235 236$ix = Bix (e−| Bi | ); $i = Bi (e−| Bi | )237 238(4)239 240The weight of local stimuli Zi is expressed by using Equation (5):241Zi =242 243q244 245y246 247($ix )2 + ($i )2248 249(5)250 2512.3. Focus Maps252A multiscale sliding window technique is applied to acquire diverse focus maps from activity253maps Zi . Two sliding windows are selected for the generation of focus maps. Firstly, a 9 × 9 window is254initialized by setting k = 9, l = 9 and σ = 1 in Equation (6). The activity maps are divided into blocks255of 9 × 9 pixels by using spatial domain filters, as in Equations (6) and (7):256G´i,σ (m, n) =257 258σ×k259 260∑261 262σ×l263 264∑265 266w(q1 , q2 ) Zi (m + q1 , n + q2 )267 268(6)269 270q1 =−σ ×k q2 =−σ ×l271 272∑273 274Gi,σ (s, t) =275 276G´i,σ (m, n)277 278(7)279 280(m,n) e Ω281 282The activity of each block is stored in the form of map scores. Furthermore, the sum of intensity283levels in each near (G1,σ=1 (s, t)) and far focus block (G2,σ=1 (s, t)) are calculated and compared with284one another to update the score maps (ζ i,σ=1 ), as given in Equations (8) and (9).285(286ζ 1,σ (m, n) =287 2881,289 290if G1,σ (s, t) > G2,σ (s, t)291 2920,293 294Otherwise295 296ζ 2,σ (m, n) = 1 − ζ 1,σ (m, n)297 298(8)299 300(9)301 302Similarly, 27 × 27 block of pixels are generated by setting k = 9, l = 9 and σ = 3 in Equation (6).303The activity maps in each near (G1,σ=3 (s, t)) and far focus block (G2,σ=3 (s, t)) are calculated and304compared with one another to update the score maps (ζ i,σ=3 ), as in Equations (8) and (9).305These multiple sliding windows result in multiple near and far focus maps. This multiscale306sliding window technique reduces the blocking artifacts in the coarse decision maps. Each map offers307different characteristic information, which plays a key role in improving the focus maps and the308fused image. These multiscale windows extract the information from original images at different309scales. It is noted that this approach has demonstrated better visual quality than the existing methods.310Each scale offers different information for image fusion, for example, a small window size focuses311on local intensity variations, whereas a large size window size extracts global variations in an image.312The information from these multiscale near-focus (ζ 1,σ=1 and ζ 1,σ=3 ) and far-focus maps (ζ 2,σ=1 and313ζ 2,σ=3 ) are combined together to form a single near-focus (D1 ) and far-focus (D2 ) map, respectively,314carrying the attributes of both scales, as in Equation (10).315AND316 317Di (m, n) ←−−−−− ζ i,σ=1 , ζ i,σ=3318 319(10)320 321After obtaining the focus maps, the next step is to generate a trimap that segments the given322images into the three different regions, i.e., focused, definite defocused, and unknown. Pixels from the323focused region have greater focus value than pixels in the defocused region [27]. The trimap T of A1 is324processed by using D1 , D2 as in Equation (11).325Tri Map326 327T ←−−−−−− Di328 329(11)330 331 Electronics 2020, 9, 472332 3336 of 16334 335In a given image I, the image matting considers it a composite of foreground I Fore and background336Each pixel is assumed to be a linear combination of I Fore and I back . Let α denote the pixel337foreground opacities then an image I can be represented as,338I back .339 340Ii = αi IiFore + (1 − αi ) Iiback341 342(12)343 344In [28], the quadratic cost function for α is derived as,345J (α) = α T Lα346 347(13)348 349where, L is defined as a matting laplacian matrix of N × N dimension.350The L is a symmetric positive definite matrix and is defined in [28] as L = H - W, where, H is a351diagonal matrix and W is a symmetric matrix. The neighborhood WM is given as,352K353 354∑355 356WM (i, j) =357 358k/(i,j)ewk359 3601361|362w363k|364=0365 366367 368ε3691 + (χi − φk )(νk +370Γ)−1 (χ j − φk )371| wk |372 373374(14)375 376where, |wk | denotes the number of pixels in the window, φk and νk represents mean and variance of377intensities in the window wk , respectively. χ represents the pixel color, e is a regularization parameter378and Γ is an identity matrix.379Finally, the obtained alpha matte α from the source images and trimap is same as the focused380region of Ii is constructed as in Equation (15).381IF (m, n) = α(m, n) I1 (m, n) + (1 − α(m, n)) I2 (m, n)382 383(15)384 3853. Results and Discussion386To show the superiority of the proposed MSIM, a comparison was performed with discrete wavelet387transform (DWT) [29], guided filtering based fusion (GFF) [13], discrete cosine transform (DCT) [30],388dense sift (DSIFT) [20], multi-scale morphological focus-measure (MSMFM) [9], and convolutional389neural network (CNN) [24] on a multifocus image dataset [31]. The proposed method was evaluated390by performing both subjective and objective assessments. These algorithms were tested on a Acer391laptop Intel(R) CoreTM i7 2.6GHz processor with 12GB RAM under a Matlab R2018b environment.392All the algorithms were executed by using the original codes made available by the authors.3933.1. Comparison of Image Matting Result394Generally, an unsupervised trimap produces better results than the supervised ones. Hence,395in practice, user specified trimaps are often necessary to achieve the high quality matting results;396however, the making of a user supervised trimap takes time, skills, and is not available for all kind of397images. In this paper, two image matting techniques have been proposed, i.e., focus maps matting and398feature based matting. The results of the proposed method are compared with feature based matting399and the closed form matting [28]. It is clearly observed that the proposed matting produces better400results compared to the existing technique (Figure 3).401 402 Electronics 2020, 9, 472403 4047 of 16405 406(a)407Source image I1408 409(b) MFM410 411(c) CFM412 413(d)414Proposed MSIM415 416(e) MFM417 418(f) CFM419 420(g)421Proposed MSIM422 423Figure 3. Results of Trimap and Alpha matte on flower image (b–d) trimaps of (a), (e–g) alpha mattes424of Figure 3a.425 4263.2. Comparison of Image Fusion with Other Methods427The proposed technique is tested on gray scale, color, and dynamic images. Figure 4 shows the428results of the proposed MSIM for “Lab” image. The source near and far focus inputs are presented in429Figure 4a,b, respectively. The fused results produced by other methods and the proposed technique430are given in Figure 4c–i. To further investigate the effectiveness, the difference of the near-focus431image with the fused images is shown in Figure 4j–p. The close up views enclosed by red and yellow432boxes are also shown at the bottom of their respective difference image. It is noted that the DWT,433DCT, and DSIFT methods produce poor edge information and contain artifacts (as shown in the434close-ups). Furthermore, GFF, MSMFM, and CNN methods also provide limited information of the435focused regions as compared to the proposed MSIM technique. Similarly, Figure 5 illustrates the436results produced by several existing and proposed algorithms for “Globe” images. To further analyze437the results, close-up views of important regions are placed at the bottom of each difference image.438In this image, the boundary region of the hand is difficult to detect since it lies on the focus transition439point. The results of fusion by other techniques in Figure 5j–o show the distorted regions and lack of440sharpness in the highlighted region. However, the proposed MSIM method has successfully fused441the complementary information from both the images, as shown in Figure 5p. It is very important442to evaluate the results of different algorithms on the color dataset shown in Figures 6a,b and 7a,b.443The outcomes of the existing techniques and the proposed method on “Flower” and “Boy” are shown444in Figures 6c–i and 7c–i, respectively. The difference between the fused and out of focus source images445is illustrated in Figures 6j–p and 7j–p, respectively. It is noted that in both the flower and boy images,446 447 Electronics 2020, 9, 472448 4498 of 16450 451the existing techniques are unable to mitigate the artifacts and blur in the focus transition area (as noted452in the close-ups of difference images). The proposed MSIM is able to preserve contrast and details453using the edge feature and multi scale image matting technique.454 455(a)456Near focus image457 458(c) DWT459 460(d) GFF461 462(e) DCT463 464(k) GFF465 466(l) DCT467 468(o) CNN469 470(f) DSIFT471 472(g)473MSMFM474 475(i)476Proposed MSIM477 478(h) CNN479 480(j) DWT481 482(b)483Far focus image484 485(m)486DSIFT487 488(n)489MSMFM490 491(p)492Proposed MSIM493 494Figure 4. Results of image fusion and their difference images on the “Lab” source images. (c–i) Fused495images obtained through fusion schemes. (j–p) Difference images obtained from the fusion results and496Figure 4b.497 498 Electronics 2020, 9, 472499 5009 of 16501 502(a)503Near focus image504 505(c)506DWT507 508(d) GFF509 510(e) DCT511 512(h)513CNN514 515(j) DWT516 517(k) GFF518 519(f)520DSIFT521 522(g)523MSMFM524 525(i)526Proposed MSIM527 528(l) DCT529 530(o)531CNN532 533(b)534Far focus image535 536(m)537DSIFT538 539(n)540MSMFM541 542(p)543Proposed MSIM544 545Figure 5. Results of image fusion and their difference images on the “Globe” source images. (c–i) Fused546images obtained through fusion schemes. (j–p) Difference images obtained from the fusion results and547Figure 5b.548 549 Electronics 2020, 9, 472550 55110 of 16552 553(a)554Near focus image555 556(c) DWT557 558(d) GFF559 560(e) DCT561 562(k) GFF563 564(l) DCT565 566(o) CNN567 568(f) DSIFT569 570(g)571MSMFM572 573(i)574Proposed MSIM575 576(h) CNN577 578(j) DWT579 580(b)581Far focus image582 583(m)584DSIFT585 586(n)587MSMFM588 589(p)590Proposed MSIM591 592Figure 6. Results of image fusion and their difference images on the “Flower” source images. (c–i)593Fused images obtained through fusion schemes. (j–p) Difference images obtained from the fusion594results and Figure 6b.595 596 Electronics 2020, 9, 472597 59811 of 16599 600(a)601Near focus image602 603(c)604DWT605 606(d) GFF607 608(e) DCT609 610(h)611CNN612 613(j) DWT614 615(k) GFF616 617(f)618DSIFT619 620(g)621MSMFM622 623(i)624Proposed MSIM625 626(l) DCT627 628(o)629CNN630 631(b)632Far focus image633 634(m)635DSIFT636 637(n)638MSMFM639 640(p)641Proposed MSIM642 643Figure 7. Results of image fusion and their difference images on the “Boy” source images. (c–i) Fused644images obtained through fusion schemes. (j–p) Difference images obtained from the fusion results and645Figure 7a.646 647 Electronics 2020, 9, 472648 64912 of 16650 651Another challenge for multi-focus fusion includes the performance in dynamic scenes.652The scenario occurs either due to the movement of the camera or the motion of the object. So it653is important to verify the effectiveness of the MSIM result with the existing ones on such scenes.654Figure 8a,b shows near and far focus “Girl” images, respectively. The results of MSIM and existing655techniques are shown in Figure 8c–i, while Figure 8j–p shows difference images. As shown in the red656and yellow boxes, the DWT, GFF, DCT, DSIFT, and MSMFM methods are unable to completely fuse657the focus regions. Moreover, the CNN has produced erosion in the fused image, whereas the proposed658MSIM has successfully mitigated the inconsistencies and limitations of the existing techniques.659 660(a)661Near focus image662 663(c) DWT664 665(d) GFF666 667(e) DCT668 669(k) GFF670 671(f) DSIFT672 673(g)674MSMFM675 676(i)677Proposed MSIM678 679(h) CNN680 681(j) DWT682 683(b)684Far focus image685 686(l) DCT687Figure 8. Cont.688 689(m)690DSIFT691 692(n)693MSMFM694 695 Electronics 2020, 9, 472696 69713 of 16698 699(o) CNN700 701(p)702Proposed MSIM703 704Figure 8. Results of image fusion and their difference images on the “Girl” source images. (c–i) Fused705images obtained through fusion schemes. (j–p) Difference images obtained from the fusion results and706Figure 8b.707 708It is observed from these visualizations that the existing methods produce artifacts, erosion,709halo effects and are unable to produce sharp boundaries of the near and far focus images. Note that710the MSIM technique not only perfectly identifies the near and far focus regions but also fuses the711complementary information in an effective manner.7123.3. Objective Evaluation Metrics713After evaluating the visual quality and quantitative assessment of different methods, it can be714clearly observed that MSIM produces a visually pleasant and high quality fusion result in almost all715cases and outperformed the existing fusion methods for multi-focus images. Five most commonly716used metrics are evaluated, i.e., Mutual Information (MI) [32], Spatial Structural Similarity (SSS)717Q AB/F [33], Feature Mutual Information (FMI) [34], Entropy (EN) [35], and Visual Information Fidelity718(VIF) [36] to verify the superiority and effectiveness of the proposed MSIM method. Table 1 shows that719the proposed MSIM gives better objective assessment results than the existing methods. Although,720the results of existing techniques are comparable in some cases (Flower and Boy); however, the metric721values obtained using the proposed MSIM generally outperforms the existing techniques.7223.4. Comparison of Computational Efficiency723In this section, the computational efficiency of different fusion methods is compared.724The execution time of these schemes for different images is shown in Table 1. The results show725that the proposed MSIM, DSIFT, and GFF consume less time as compared to the other algorithms DCT,726DWT, MSMFM, and CNN. The MSMFM algorithm uses a multi-scale morphological gradient based727feature, therefore taking longer processing time than DSIFT. Whereas, GFF integrates the source images728by using a global weight based scheme; however, it still takes less computation time and produces729satisfactory results.730The proposed method utilizes the contrast enhancement, SSGSM based edge extraction, sliding731window based local and global operations to create activity maps and trimap. The sliding window732method, activity maps generation, their comparison, and a trimap generation are time consuming733tasks. Although, the proposed algorithm consumes more processing time as compared to the existing734ones; however, it produces the best unsupervised image matting and image fusion results.735 736 Electronics 2020, 9, 472737 73814 of 16739 740Table 1. The quantitative assessment of different fusion methods.741Images742 743Lab744 745Globe746 747Flower748 749Boy750 751Girl752 753Fusion Methods754 755MI [32]756 757Q AB/F [33]758 759FMI [34]760 761EN [35]762 763VIF [36]764 765Time/s766 767DWT [29]768 7698.2152770 7710.7239772 7730.8190774 7757.0474776 7770.9138778 7794.02780 781GFF [13]782 7837.9114784 7850.7279786 7870.8191788 7897.0602790 7910.9149792 7933.11794 795DCT [30]796 7978.5263798 7990.7460800 8010.9197802 8036.9819804 8050.9143806 80711.56808 809DSIFT [20]810 8118.5212812 8130.7478814 8150.9097816 8177.0759818 8190.9171820 8216.65822 823MSMFM [9]824 8258.7995826 8270.6864828 8290.9196830 8316.9885832 8330.9161834 8355.79836 837CNN [24]838 8398.6812840 8410.7471842 8430.9196844 8456.9974846 8470.9159848 8497.88850 851Proposed852 8538.8322854 8550.7474856 8570.9386858 8597.1759860 8610.9980862 8635.08864 865DWT [29]866 8678.1910868 8690.7246870 8710.8892872 8737.7037874 8750.9240876 87711.09878 879GFF [13]880 8818.7664882 8830.7726884 8850.8935886 8877.7412888 8890.9476890 8919.78892 893DCT [30]894 8959.1845896 8970.7731898 8990.8939900 9017.6990902 9030.9374904 90510.98906 907DSIFT [20]908 9099.1435910 9110.7746912 9130.8938914 9157.6989916 9170.9437918 9195.51920 921MSMFM [9]922 9239.3739924 9250.7711926 9270.8940928 9297.7389930 9310.9439932 9336.55934 935CNN [24]936 9379.2397938 9390.7701940 9410.8927942 9437.7458944 9450.9472946 9478.01948 949Proposed950 9519.4316952 9530.7733954 9550.8943956 9577.7479958 9590.9480960 9617.06962 963DWT [29]964 9655.6452966 9670.6536968 9690.8773970 9717.1701972 9730.9064974 97516.93976 977GFF [13]978 9797.3290980 9810.6944982 9830.8908984 9857.1915986 9870.9200988 98910.13990 991DCT [30]992 9937.8561994 9950.6785996 9970.8861998 9997.43311000 10010.92631002 100312.031004 1005DSIFT [20]1006 10078.00571008 10090.69471010 10110.88571012 10137.43161014 10150.93041016 10174.611018 1019MSMFM [9]1020 10217.92331022 10230.69301024 10250.89151026 10277.18731028 10290.91401030 10315.981032 1033CNN [24]1034 10353.07731036 10370.69511038 10390.89121040 10417.18721042 10430.91771044 10457.321046 1047Proposed1048 10498.14581050 10510.79401052 10530.89361054 10557.58971056 10570.93671058 10596.911060 1061DWT [29]1062 10637.53211064 10650.72061066 10670.88141068 10697.53711070 10710.89351072 10739.951074 1075GFF [13]1076 10777.63161078 10790.74481080 10810.87171082 10837.53101084 10850.80971086 10875.971088 1089DCT [30]1090 10918.08521092 10930.74091094 10950.87141096 10977.56691098 10990.90351100 110111.081102 1103DSIFT [20]1104 11058.17651106 11070.74371108 11090.87211110 11117.53881112 11130.90481114 11153.801116 1117MSMFM [9]1118 11198.20811120 11210.74181122 11230.87171124 11257.24021126 11270.90261128 11297.551130 1131CNN [24]1132 11332.99661134 11350.74661136 11370.88191138 11397.53861140 11410.90721142 11438.011144 1145Proposed1146 11478.29611148 11490.74871150 11510.88261152 11537.56731154 11550.90771156 11577.751158 1159DWT [29]1160 11615.62261162 11630.59391164 11650.81891166 11677.84771168 11690.64861170 11715.391172 1173GFF [13]1174 11758.08201176 11770.69021178 11790.82131180 11817.84751182 11830.64481184 11855.881186 1187DCT [30]1188 11898.72221190 11910.67881192 11930.81681194 11957.82261196 11970.73321198 11996.441200