- Research Article
- Open Access
Automatic Image Interpolation Using Homography
© The Author(s) 2010
- Received: 8 July 2009
- Accepted: 13 April 2010
- Published: 20 May 2010
While taking photographs, we often face the problem that unwanted foreground objects (e.g., vehicles, signs, and pedestrians) occlude the main subject(s). We propose to apply image interpolation (also known as inpainting) techniques to remove unwanted objects in the photographs and to automatically patch the vacancy after the unwanted objects are removed. When given only a single image, if the information loss after the unwanted objects in images being removed is too great, the patching results are usually unsatisfactory. The proposed inpainting techniques employ the homographic constraints in geometry to incorporate multiple images taken from different viewpoints. Our experiment results showed that the proposed techniques could effectively reduce process in searching for potential patches from multiple input images and decide the best patches for the missing regions.
- Multiple Image
- Foreground Object
- Point Pair
- Image Inpainting
- Epipolar Constraint
To make the 3D architecture models more visually realistic, 2D images are sometimes used as texture patches. But when photographs are taken, foreground objects (e.g., tourists, vehicles, signs, etc.) sometimes occlude the main subjects. Image inpainting is the process to remove unwanted objects in the photographs and to patch the vacancy.
In relevant image inpainting literature, to remove damaged or undesired objects in the image, the most common method is to employ image editing tools to manually select the unwanted objects and then filled the target regions with pre-selected color. An example is shown in white in Figure 3(b). The process to patch the vacancy after unwanted objects are removed is commonly referred as image inpainting and texture synthesis in the literature [8–11].
The inpainting algorithms proposed by Oliverira et al.  in 2001 and the Fast Marching Method (FMM) proposed by Telea  in 2004 improve the speed while patching small missing or damaged regions but cause the image blurred when target regions inpainted are large.
In 2004, Criminisi et al.  combined the advantages of texture synthesis and image inpainting for large objects removal and inpainting. In 2005, Cheng et al. proposed a robust algorithm  with improved priority computation in . Sun et al. (2005) proposed to first use image structure propagation and then fill the target regions .
In single image inpainting, the patching process relies only on the remaining image areas after undesired objects are removed. When the image information loss is too large after object removal, the patching results are usually undesirable. Some previous works also proposed the use of multiple views for inpainting [13–15]. However, the method proposed in  relies mainly on landmarks matching without using any geometric constraints. In some cases, manual identification may still be required. In , the input is a series of images taken from a moving camera and motion-based background selection is employed, while our proposed method takes arbitrary views and do not reply on motion information. The proposed method in  requires multiple calibrated views where our proposed method can use uncalibrated views for inpainting. Other applications of image inpainting are to restore old films and to remove and edit image objects automatically [16, 17].
Previous research used an image for image inpainting. In this paper, we incorporate multiple images taken from different viewpoints for image inpainting. Our idea is that the regions needed for image inpainting can be correctly filled by other images taken from different viewpoints. However, incorporating multiple images taken from different viewpoints creates a challenging problem: automatic point correspondence among multiple images taken from different viewpoints is needed. In this paper, we first apply the homography property to solve the point correspondence problem among multiple images taken from different viewpoints for image inpainting. Based on the homography property, we used a robust method called the Least Median of Squares (LMedS) to achieve correct point correspondences. Our main contribution is to propose an automatic image interpolation algorithm for image inpainting.
The idea of using multiple images from different shooting angles is to try to recover objects that may not be occluded in all shootings. In this following section we propose to use the homography and rectification for image inpainting with multiple images.
In this section we discuss the image geometric characteristic we propose to employ in image inpainting with multiple images. Projective geometry refers to the relationship among images that are formed by the projections of the light reflection of objects in 3D space into 2D images, taken by different camera rotation and translation. We then discuss in details how to employ multiple source images in image inpainting.
3.1. Camera Geometry and Camera Model
where is 3 4 projection matrix describing the perspective projection process, and are vectors containing the homogeneous coordinates of the world points, and is a scale factor, respectively, image points.
where and represent the focal length divided by the horizontal and vertical pixel dimensions, is a measure of the skew, and is the principal point.
3.2. Two-View Geometry
Consider the image points and of a 3D point observed by two cameras with optical centers and . These five points form a common plane, that is, defined as the so-called epipolar plane. The points and are called the epipoles of the two cameras. The epipole is the projection of the optical center of the first camera in the image observed by the second camera and vise versa. If and are images of the same point, then must lie on the epipolar line associated with , that is so-called the epipolar constraint.
The epipolar constraint plays an important role in stereo vision analysis. When the internal camera parameters are known, the epipolar constraint can be represented algebraically by a 3 3 matrix, called the essential matrix. Otherwise, the epipolar constraint represented by a 3 3 matrix is called the fundamental matrix, F.
where F is the 3 3 fundamental matrix.
If we have n matched point pairs from the same 3D plane, the above equation can be used to solve the 3 3 homography matrix H by applying the SVD method.
In Section 2 we present the image inpainting results of different techniques with a single image. But for complex scenes, we show that after foreground removal, the information loss is too great that the missing regions cannot be recovered from the remaining image areas. The idea of using multiple images from different shooting angles is to try to recover objects that may not be occluded in all shootings. In this section we discuss how to apply homography and rectification for image inpainting with multiple images.
Because there exist many flat surfaces where on architecture and there is a certain geometry relationship called homorgraphy, between two images, we can exploit this characteristic to locate the best fitted image patches for image inpainting.
To calculate the homography matrix H, one usually select four or more corresponding point sets manually. In the paper, we employ the automatic process to select potentially better corresponding point sets  to compute the homography matrix H instead of manually selecting corresponding point pairs.
First, within the selected area, we use the Sum of Squared Differences (SSD) method to locate the most likely corresponding point pairs in the two input images. The point pair with the least SSD value within there surrounding window is considered the potential corresponding point pair. The intermediate results of this step are as shown in Figure 12.
However, the SSD value can still produce erroneous corresponding point pairs because of conditions such as lighting and occlusion. We then employ the Least Median of Squares (LMedS) method to exclude the use of the corresponding point pairs to compute the homography matrix H.
A random sampling strategy similar to RANSAC is adopted because the median is not differentiable. Instead of using the consensus of all data points, sample of size is randomly selected and the corresponding homography matrix is computed. The residual errors of all data points with respect to the homography matrix are computed and sorted in a table as shown in Figure 13. The model with the least median (minimum median residual errors) is chosen. The LMedS method can tolerate up to 50% of outliers; that is, without changing the objective function value, the LMedS method can have up to half of the data points arbitrarily far from the true estimate.
The proposed automatic process is stated as follows.
(1)Determine the features points in images with the Harris Corner Detector method (as shown in Figure 11).
(2)Select a region in the image panes. Within the selected area, use the Sum of Squared Differences (SSD) method to locate the most likely corresponding point pairs in the two input images. The intermediate results of this step are as shown in Figure 12.
The complete algorithm of image inpainting with homography is detailed as follows.
(1)Take sequential images of the objects or scenes with different shooting angles and extract the image frames, .
(2)Compute the homography geometry relationship among .
(3)Use the homography matrix H to transform the source images.
(4)Select manually the target region.
(5)Inpaint the target region automatically:
(a)get the contour of the target region to prioritize the patching order. Compute the target patch and the source patches;
(b)use SSD to compute the similarity of the target patch and the source patches;
(c)fill the target patch with the most similar source patch.
(6)Update the target area.
(7)Repeat steps 5~6 until the entire target area is patched.
Traditional image inpainting techniques employ a single input image. The patching process relies only on the remaining image areas after the undesired objects are removed. When the image information loss is too great after object removal, the patching results are usually undesirable.
We propose inpainting techniques that employ multiple images from different viewpoints. From multiple source images we can extract image patches that are not occluded in some images. The proposed inpainting techniques employ the homographic constraints in geometry among image frames from multiview images to assist the inpainting process. Our experiment results support that the proposed method can reduce the search process and increase the accuracy in inpainting.
Because we use multiple source images taken from different angles and under different lighting conditions, the patched area may suffer slightly inconsistency in terms of brightness. It is suggested to explore solutions to overcome this problem with increased image resolution in the future and investigate other geometric constraints that can be applied on the image inpainting process.
This work was partially supported by the National Science Council, Taiwan, under the Grant no. NSC98-2631-H-211-001 and NSC97-2221-E-011-090.
- Oliveira MM, Bowen B, McKenna R, Chang Y-S: Fast digital image inpainting. Proceedings of the International Conference on Visualization, Imaging and Image Processing(VIIP '01), 2001, Marbella, SpainGoogle Scholar
- Telea A: An image inpainting technique based on the fast marching method. Journal of Graphics Tools 2004, 9(1):25-36.View ArticleGoogle Scholar
- Criminisi A, Pérez P, Toyama K: Region filling and object removal by exemplar-based image inpainting. IEEE Transactions on Image Processing 2004, 13(9):1200-1212. 10.1109/TIP.2004.833105View ArticleGoogle Scholar
- Cheng W-H, Hsieh C-W, Lin S-K, Wang C-W, Wu J-L: Robust algorithm for exemplar-based image inpainting. Proceedings of the International Conference on Computer Graphics, Imaging and Visualization, 2005Google Scholar
- Drori I, Cohen-Or D, Yeshurun H: Fragment-based image completion. Proceedings of the ACM SIGGRAPH Conference on Computer Graphics, 2003Google Scholar
- Cheng K-Y: Research on improving exemplar-based inpainting, M.S. thesis. Department of Computer Science and Information Engineering, National Taiwan University; 2005.Google Scholar
- Levin A, Zomet A, Weiss Y: Learning how to inpaint from global image statistics. Proceedings of the 9th IEEE International Conference on Computer Vision, October 2003, Nice, France 305-312.View ArticleGoogle Scholar
- Efros AA, Leung TK: Texture synthesis by non-parametric sampling. Proceedings of the 7th IEEE International Conference on Computer Vision (ICCV '99), September 1999, Corfu, Greece 1033-1038.View ArticleGoogle Scholar
- Efros AA, Freeman WT: Image quilting for texture synthesis and transfer. Proceedings of the ACM SIGGRAPH Conference on Computer Graphics, 2001 341-346.Google Scholar
- Jia J, Tang C-K: Image repairing: robust image synthesis by adaptive ND tensor voting. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, June 2003 643-650.Google Scholar
- Tang P: Application of non-parametric texture synthesis to image, M.S. thesis. Computer Science, University of New Mexico; 2004.Google Scholar
- Sun J, Yuan L, Jia J, Shum H-Y: Image completion with structure propagation. Proceedings of the ACM SIGGRAPH Conference on Computer Graphics, August 2005 861-868.Google Scholar
- Kang SH, Chan TF, Soatto S: Inpainting from multiple views. Proceedings of the 1st International Symposium on 3D Data Processing Visualization and Transmission (3DPVT '02), 2002Google Scholar
- Rasmussen C, Korah T: Spatiotemporal inpainting for recovering texture maps of partially occluded building facades. Proceedings of the IEEE International Conference on Image Processing (ICIP '05), September 2005, Genoa, Italy 125-128.Google Scholar
- Zokai S, Esteve J, Genc Y, Navab N: Multiview paraperspective projection model for diminished reality. Proceedings of the 2nd IEEE/ACM International Symposium on Mixed and Augmented Reality, 2003 217-226.Google Scholar
- Patwardhan KA, Sapiro G, Benalmio M: Video inpainting of occluding and occluded objects. Proceedings of the IEEE International Conference on Image Processing (ICIP '05), September 2005, Genoa, Italy 69-72.Google Scholar
- Wexler Y, Shechtman E, Irani M: Space-time video completion. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR '04), July 2004 1: 120-127.Google Scholar
- Tang C-Y, Chou H-L, Kuo W-H, Hu P-C: Using 3D information to establish stereo correspondence from uncalibrated images. Proceedings of the 18th IPPR Conference on Computer Vision, Graphics and Image Processing, 2005, Taipei, TaiwanGoogle Scholar
- Tang C-Y, Chou H-L, Ding Y-H: Robust methods to estimate fundamental matrices using planar constraint. Proceedings of the IEEE International Conference on Systems, Man and Cybernetics (SMC '06), 2006, Taipei, TaiwanGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.