AUTOMATIC IMAGE RETARGETING USING SALIENCY BASED MESH PARAMETERIZATION

Transcription

1 S.Sai Kumar et a. / (IJCSIT Internationa Journa of Computer Science and Information Technoogies, Vo. 1 (4, 010, AUTOMATIC IMAGE RETARGETING USING SALIENCY BASED MESH PARAMETERIZATION 1 S.Sai Kumar, A.Sudhir Babu Department of Computer Science Engineering PVP Siddhartha Institute of Technoogy, Vijayawada, Andhra Pradesh, India. 1 E-mai:saiumar_vij@yahoo.com E-mai:asbabu@hotmai.com Abstract Automatic image retargeting is used for arge image.that are to be fit in sma size dispay devices. Without any oss of information.our proposed methods one is saiency based mesh parameterization method is used to retarget the image. Stretch-Based Mesh Parameterization is generating on saiency image. Graph- Based visua saiency is used to easiy find out the objects in the image. Mesh Generation on that saiency object part it shoud be retarget the image. Another method is comparison of seam-carving and Automatic retargeting method. In this we show the same image as differentiate by axes to separate one is oss of information and structure of the image is changed is caed seam-carving. Another one is without oss of information and shoud not change the structure of the image is caed retargeted method. Key words:automatic image retargeting, Mesh parameteriztion,seam-carving. I. INTRODUCTION Now a days mobie devices ie ceuar phones,pda s are increasing more and more.these are a portabe dispay image functionaities. Mobie devices having sma screens[1] with different aspect ratio and much more highest resoutions.those are to be fit in target dispay. The probem of adapting the images to fit in various trget screens by using image retargeting. A common functioaity of image retrgeting is to rescae the origina image according to the target screen size. Scaing method has the probem. That is some important portion of the image shoud be decreased. So visua distortion is introduced. The next method is cropping to retarget the image. It is used to crop the important area of the image. The main disadvantages of cropping are contextua content of the particuar image to be re moved. Another method is seamcarving [4] [6]. In this image resizes by iterativey remove the ow resoution pixe vaues. The main drawbac of seamcarving is to brea the objects of the image. So image structure is changed. The ast method for image resize by using mesh parameterization [8]. In this mesh is generation on coor images. Where we seect the objects from the image is manuay by using bue edge. Reduce the edges of the objects and to fit in target size. It aso has the disadvantage that is mesh generation on coor image is more time taing processes. To find out the threshod vaues of the image it taes ot of time. Another one is by seecting the objects of the image is manuay. To overcome these probems our method is introduced that is Automatic image retargeting [3] using saiency [] based mesh parameterization. In this we define two areas one is saiency based mesh parameterization and another one is comparison of seam-carving and retargeting methods. Now our proposed method Initiay taen the input coor image. Input image is discriminated by row-coumn wise. Cacuate the edges by using CANNY EDGE DETECTION agorithm. Appy saiency to change the coor image to Gray- Scae image. Where the objects are found it taes high resoution and it indicates white coor, ow resoution indicates bac coor. The saiency image is to be discriminated by rowcoumn wise. Where the high resoution is found it indicates bac discrimination points, ow resoution indicates white discrimination points. Appy Stretch-Based Mesh Parameterization and using Deaunay Trianguation to retarget the image. With out any oss of information. To produce the retarget image. Comparison of seam-carving and automatic retargeting methods. II.RELATED WORK The probem of retargeting images to sma screens has been considered by many researchers. Amost a image browsers on sma-screen devices provide scaing functionaity. A arge 73

2 S.Sai Kumar et a. / (IJCSIT Internationa Journa of Computer Science and Information Technoogies, Vo. 1 (4, 010, image is uniformy down samped to fit the target screens. Important objects in the image are often sma and difficut to identify. To account for different aspect ratios, either the distortion caused by aspect ratio change is introduced or bac etter box is used to fi in the ban space, thus wasting the precious space. Many methods identify an important region in the image and crop this region to fit target screen. These methods usuay rey on image anaysis, such as attention mode extraction and object detection, to identify the important region. Cropping-based retargeting methods can effectivey preserve important information of the origina image, at the expense of osing vauabe context information. A fish-eye view warping scheme to achieve the goa of emphasizing the important region whie retaining the surrounding context. This method can effectivey achieve the focus + context view of the origina image. However, it creates noticeabe distortion, especiay in the contextua region. Seam-carving method for resizing images. Their method iterativey carves unnoticeabe seams to reduce the image size. However, it often breas image structures since the carving procedure ony reies on ow-eve saiency information and fais to consider image structures. Both Setur et a. and Cheng et a. proposed context re composition for preserving saient information. These methods first segment saient objects based on the detected saiency prior. Then they fi in the hoes eft after object segmentation. Finay, they compose saient objects bac onto the resized bacground video frames based on a set of aesthetic criteria. II. RETARGETING SALIENCY BASED MESH PARAMETERIZATION Image retargeting attempts to adapt a source image to dispay of smaer size [1]. And different aspect ratio than originay intended. We formuate image retargeting as a mesh parameterization probem that emphasizes the important content whie retaining the surrounding bacground with sight distortion. To emphasize the input image is taen, then appy the discrimination of the input image and cacuate the edges of the input image (Section III-A. To emphasize saient objects in image, we first cacuate a saiency map of the source image and appy the discrimination of the saiency map image (Section III-B. Then we buid a controing mesh of the saiency map image that is consistent with the underying image structures, and associate the saiency information with the source mesh (Section III-C. In this way, retargeting is transformed into a parameterization probem of finding a target mesh with the desired resoution. In order to improve visua quaity of the target image, we interpret preservation of saiency and image structure information as constraints of parameterization.afterwards, to generate target size of saiency image is soved by a constrained stretch-based mesh parameterization scheme. Automatic image retargeting agorithm to generate retargets image size. With out any information oss and aso compare the seam-carving and image retargeting (Section III -D. A. Importance Computation Given Input Image Fig : 1 Indicates given input Image. Input Image After Discretization Fig : Indicates input Image after Discritization. Fow-Chart of Saiency Based Mesh Parameterization Fow-Chart 1: Indicates the Reprasentation of Agoritham. 74

3 S.Sai Kumar et a. / (IJCSIT Internationa Journa of Computer Science and Information Technoogies, Vo. 1 (4, 010, In this section we can read the input image, and that can be discritized by using row wise- coumn wise discretization. In this Discretization to equa portioning points is generated by rows and coumns. It is used to easiy identify the object of the image. After that we find out the edges of the image by using Canny Edge Detection Agorithm. Fig: 3 Indicates to find the Edges. Canny Edge Detection Agorithm EDGE IMAGE The Canny agorithm uses an optima edge detector based on a set of criteria which incude finding the most edges by minimizing the error rate, maring edges as cosey as possibe to the actua edges to maximize ocaization, and maring edges ony once when a singe edge exists for minima response [1]. According to Canny, the optima fiter that meets a three criteria above can be efficienty approximated using the first derivative of a Gaussian function equation (1, (. G (, = 1 e G(, e G(, e (1 ( The first stage invoves smoothing the image by convoving with a Gaussian fiter. This is foowed by finding the gradient of the image by feeding the smoothed image through a convoution operation with the derivative of the Gaussian in both the vertica and horizonta directions. The -D convoution operation is described in the foowing equation (3. Where: g(, = convoution erne I(x,y = origina image I (x,y = fitered image N + 1 = size of convoution erne Both the Gaussian mas and its derivative are separabe, aowing the -D convoution operation to be simpified. This optimization is not imited to software impementation ony, but appies to hardware impementation as we, as shown in the next section. The non-maxima suppression stage finds the oca maxima in the direction of the gradient, and suppresses a others, minimizing fase edges. The oca maxima are found by comparing the pixe with its neighbors aong the direction of the gradient. This heps to maintain the singe pixe thin edges before the fina threshoding stage. Instead of using a singe static threshod vaue for the entire image, the canny agorithm introduced hysteresis threshoding, which has some adaptivey to the oca content of the image. There are two threshod eves, th, high and t, ow where th > t. Pixe vaues above the th vaue are immediatey cassified as edges. By tracing the edge contour, neighboring pixes with gradient magnitude vaues ess than th can sti be mared as edges as ong as they are above t. This process aeviates probems associated with edge discontinuities by identifying strong edges, and preserving the reevant wea edges, in addition to maintaining some eve of noise suppression. Whie the resuts are desirabe, the hysteresis stage sows the overa agorithm down consideraby. The performance of the canny agorithm depends heaviy on the adjustabe parameters, σ, which is the standard deviation for the Gaussian fiter, and the threshod vaues, th and t. σ aso contros the size of the Gaussian fiter. The bigger the vaue for σ, the arger the size of the Gaussian fiter becomes. This impies more burring, necessary for noisy images, as we as detecting arger edges. As expected, however, the arger the scae of the Gaussian, the ess accurate is the ocaization of the edge. Smaer vaues of σ impy a smaer Gaussian fiter which imits the amount of burring, maintaining finer edges in the image. The user can taior the agorithm by adjusting these parameters to adapt to different environments with different noise eves. Resuts can be further improved by performing edge detection at mutipe resoutions using muti-scae representations, simiar to the Marr-Hidreth agorithm [1]. This is achieved using different standard deviations, which correspond to different resoution versions of the image. Edges have zero crossing at mutipe scae vaues. B. GBVS Saiency Constraint In this section the information of important objects in an image is necessary for image retargeting to emphasize those objects. Athough semantic image understanding is beyond the state-of-the-art, heuristic rues have been used successfuy as an aternative. Lie previous wor, we use the ow-eve saiency information as we as high-eve object information to infer important objects. Saiency has been used in many appications [] [7] to detect the important region in an image. It has been extensivey expored in the past severa years. We use the contrast-based saiency detection method proposed by Ma et a. Itti et a. have defined a saiency-based visua attention mode for scene anaysis. In this paper, we adopt the approaches in to generate the three channe saiency maps, coor contrasts, intensity contrasts, and orientation contrasts, 75

4 S.Sai Kumar et a. / (IJCSIT Internationa Journa of Computer Science and Information Technoogies, Vo. 1 (4, 010, by using the approaches proposed and then buid the fina saiency map using the iterative method proposed.as iustrated in saiency attention is determined by the number of saiency regions and their brightness, area, and position in the gray saiency map [5], as shown in Fig. 4. However, in order to reduce adaptation time we binarize the saiency map to find the regions that most iey attract human attention, i.e., using a binarization threshod on brightness, which is estimated adaptivey. Thus the saiency attention vaue is (4. AV B. W saiency (4 (, R saiency Fig: 4 Indicates Saiency Attention Detection. Discretization of saiency part of the image Fig: 7 Indicates Discritization Saiency part of the Image. The above figures indicating the input image can be converted into saiency map due to find out the objects where in this white coor indicate the objects are present in the image, and bac coor indicates bacground of the image Discrimination saiency map is the technique which is used to discritizes the objects and bacground where bac point indicates the object of the image and white points indicates the bacground of the image. C. Constrained Mesh Parameterization Where denotes the brightness of pixe point ( in the saiency region, and is the position weight of that pixe. Since peope often pay more attention to the region near the image center, a normaized Gaussian tempate centered at the image is used to assign the position weight. Now our experiment resuts of the saiency map of the image is Deaunay Trianguation Fig: 8 Indicates to construct Deaunay Trianguation mesh Fig: 5 Indicates After Appying a Saiency Map. Reconstructed Image x origina image GBVS map Fig: 9 Indicates Retargeted saiency image. most saient (75%ie parts saiency map overaid Fig: 6 Indicates saiency map overaid. Fig: 10 Indicates Retargeted saiency image 76

5 S.Sai Kumar et a. / (IJCSIT Internationa Journa of Computer Science and Information Technoogies, Vo. 1 (4, 010, The Stretch-Based Mesh Parameterization With the cacuated edge engths, M 0 can be obtained by one stretch-based parameterization method. Mesh parameterization has been extensivey studied in computer graphics [8], [9]. These methods tae different metric criterions as the objective of energy minimization process Among them. Stretch-based methods wor we when reducing the goba mesh distortion [10]. Our concern is computing one D-to-D parameterization, which is different from the aforementioned 3D-to-D process (Fig. 10. We first describe some reated notations. Let {P i = (X i, Y i I=1,..,n} denote the nodes of M. {S i = (U i, V i i=1,.n} Represent the corresponding nodes of M0 to be soved in texture space.m 0 has the same topoogy as M, S i Corresponds to P i, and edge (S i, S j corresponds to (P i, P j. The 3D ength ij of (S i, S j is obtained using the aforementioned method. M 0 is competey characterized by the engths of its edges. As each edge of M 0 has been figured out, M 0 can be computed by minimizing the foowing energy function equation (5. ( S S n E (5 t (, edges n Expoiting the symmetry of the above equation, energy gradients on point S i are in equation (6. Et u 8 Et 8 v (, edges ( S S n.( u u n ( S S n.( v v (, (6 edges n When M is dense, directy soving the above equations may cause adjacent trianges of M 0 to fip over, eading to invaid topoogy. This is caused by inverting the orientation of the three points of triange. To tace this issue, we revise the energy function by penaizing the orientations of trianges using the sgn function. Assume that the adjacent trianges incident upon edge (Si, Sj are TS1 = T (SiSjS and TS = T(SiSjS. Their corresponding trianges in the input image are TP1 = T(PiP1Pj, and TP = T(PiPjP. For each pair of corresponding trianges, the orientations of points shoud be equa. W To achieve this, we define the equation (7. ij det( S sgn S min(det(, S S S S 1.det( P P, S S i 1 1, P P j. det( P P i 1, P P (7 The energy function is then transformed into equation (8. ( wij. S S n Et (, edges (8 n Where the coefficient w ij penaizes the triange in M 0.whose orientation of points fips over with respect to its corresponding Triange in M. If so, w ij is chosen as _1, otherwise, þ1. With this energy function, a vaid mesh in texture space is obtained. Minima vaue of (8 is computed by the mutidimensiona Newton s method. For each iteration of Newton s method, one mutigrain sover is adopted for soving the sparse inear equations. In practice, it converges to the fina soution within severa seconds. Once M 0 is obtained with the parameterization process, a new texture can be mapped onto the ROI of the input image. Since the parameterization taes into account the underying geometry of ROI, the new texture deforms naturay with respect to the underying surface. Our experimenta resuts demonstrate that the distortion effects of the new textures are visuay peasing. D. Automatic Image retargeting method Fow Chart of Automatic Image Retargeting method j 1 Fig: 11 The orientation of the trianges (a in image space shoud eep Consistent with that of their corresponding ones (b in retargeted texture space. Fow-Chart : Indicates the comparison of seam-carving and retargeting methods 77

6 S.Sai Kumar et a. / (IJCSIT Internationa Journa of Computer Science and Information Technoogies, Vo. 1 (4, 010, Automatic image retargeting agorithm to generate retargets image size. With out any information oss and aso compare the seam-carving and image retargeting. Appy seam-carving method [4]. In this image resizes by iterativey remove the ow resoution pixes. Seam-carving is to brea the objects of the image. So image structure is changed of images. This Seam Carving is a technique proposed by Avidan and Shamir for content-aware resizing technique beongs to the category of pixe removing. The basic idea is to remove paths of ow energy pixes (seams from top to bottom or from eft to right which are not so important for the understanding of the image content. The remova of each seam causes a reduction of the image size by one where vertica seams reduce the width and horizonta seams reduce the height. Fig: 1Left: Origina image. Center: The coor indicates the order of the seams (white seams with ow energy are removed first; red seams are preserved as ong as possibe. Right: Image after the seams have been removed. The Automatic image retargeting method is a technique is not ony for content-aware resizing beongs to the category of pixe removing and aso remove ow energy pixes but without oss of information. Seam-carving retargeting method Fig: 13 eft images indicate seam-carving and right image indicates retargeting method LIMITATION Our method retargets images with saient objects by a process of saiency-based mesh parameterization. Saient object extraction reies on image saiency and face detection. Since D image is prone to variations of body poses, face expression, and iumination, the robust human face and body detection is sti chaenging. This maes accurate saient object extraction difficut. In our test exampes, our method fais to capture some important human bodies with varied face poses or iumination. It is one drawbac of our approach. CONCLUSION AND FUTURE WORK We have proposed an effective Automatic image retargeting on saiency based mesh parameterization method based on a mesh image representation. For the first time, we formuate image retargeting as a Saiency based mesh parameterization probem that aims to find a homomorphous mesh with the desired size of the target dispay. Our method achieves emphasizing important objects whie retaining the surrounding context. Our scheme of using mesh representation for image adaptation enabes easy preservation of images structures. In future, texture mapping on retargeted image to generate coor image. REFERENCES [1] [1] L.-Q. Chen, X. Xie, X. Fan, W.-Y. Ma, H.-J. Zhang and H.-Q. Zhou, A visua attention mode for adapting images on sma dispays Mutimedia Syst., vo. 9, no. 4, pp , 003. [] V. Setur, T. Lechner, M. Nienhaus, and B. Gooch, Retargeting images and video for preserving information saiency, IEEE Computer. Graph. App., vo. 7, no. 5, pp , Sep. /Oct [3] V. Setur, S. Taagi, R. Rasar, M. Geicher, and B. Gooch, Automatic image retargeting, in Proc. 4th Int. Conf. Mobie and Ubiquitous Mutimedia, 005. [4] S. Avidan and A. Shamir, Seam carving for content-aware image resizing, ACM Trans. Graph., vo. 6, no. 3, pp , 007. [5] L. Itti, C.Koch, and E. Niebur, A mode of saiency-based visua attention for rapid scene anaysis, IEEE Trans. Pattern Ana. Mach. Inte., vo. 0, no. 11, pp , Nov [6] F. Liu and M. Geicher, Automatic image retargeting with fisheye view warping, in Proc. ACM UIST, 005, pp [7] Y. Li, Y.-F. Ma and H.-J. Zhang, Saient region detection and tracing in video, in IEEE Int. Conf. Mutimedia and Expo, 003, pp [8] Yanwen Guo, Feng Liu, Jian Shi, Zhi -Hua Zhou, Image Retargeting Using Mesh Parameterization, IEEE Transaction on Mutimedia, VOL. 11, NO. 5, EXPERIMENTAL RESULTS Seam-carving retargeting method Fig: 14 eft images indicate seam-carving and right image indicates retargeting method. A 78

7 S.Sai Kumar et a. / (IJCSIT Internationa Journa of Computer Science and Information Technoogies, Vo. 1 (4, 010, B The origina size of the image is 800*600, The above Fig A, B, C is the experimenta resuts of our paper (A resut of size 30*30, (B resuts of size 30*40, (C resuts of size 51*70 C 79