## A feature-based, robust, hierarchical algorithm for registering pairs of images of the curved human retina (2002)

### Cached

### Download Links

- [www.cs.rpi.edu]
- [www.vision.cs.rpi.edu]
- [www.ecse.rpi.edu]
- [www.cs.rpi.edu]
- DBLP

### Other Repositories/Bibliography

Venue: | IEEE Transactions on Pattern Analysis and Machine Intelligence |

Citations: | 57 - 18 self |

### BibTeX

@ARTICLE{Can02afeature-based,,

author = {Ali Can and Charles V. Stewart and Badrinath Roysam and Howard L. Tanenbaum},

title = {A feature-based, robust, hierarchical algorithm for registering pairs of images of the curved human retina},

journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},

year = {2002},

volume = {24},

pages = {347--364}

}

### OpenURL

### Abstract

AbstractÐThis paper describes a robust hierarchical algorithm for fully-automatic registration of a pair of images of the curved human retina photographed by a fundus microscope. Accurate registration is essential for mosaic synthesis, change detection, and design of computer-aided instrumentation. Central to the newalgorithm is a 12-parameter interimage transformation derived by modeling the retina as a rigid quadratic surface with unknown parameters, imaged by an uncalibrated weak perspective camera. The parameters of this model are estimated by matching vascular landmarks extracted by an algorithm that recursively traces the blood vessel structure. The parameter estimation technique, which could be generalized to other applications, is a hierarchy of models and methods: an initial match set is pruned based on a zeroth order transformation estimated as the peak of a similarity-weighted histogram; a first order, affine transformation is estimated using the reduced match set and least-median of squares; and the final, second order, 12-parameter transformation is estimated using an M-estimator initialized from the first order estimate. This hierarchy makes the algorithm robust to unmatchable image features and mismatches between features caused by large interframe motions. Before final convergence of the M-estimator, feature positions are refined and the correspondence set is enhanced using normalized sum-of-squared differences matching of regions deformed by the emerging transformation. Experiments involving 3,000 image pairs �1; 024 1; 024 pixels) from 16 different healthy eyes were performed. Starting with as low as 20 percent overlap between images, the algorithm improves its success rate exponentially and has a negligible failure rate above 67 percent overlap. The experiments also quantify the reduction in errors as the model complexities increase. Final registration errors less than a pixel are routinely achieved. The speed, accuracy, and

### Citations

2449 |
Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography
- Fischler, Bolles
- 1981
(Show Context)
Citation Context ...ferentiable, partially due to use of the median and partially due to the selection process in choosing the best correspondence for each …A; t1†. As a result, we use a random sampling search technique =-=[23]-=-, [41], [50] to find the approximate minimum of the objective function. The method differs only slightly from the techniques proposed in the original LMS [41] and Random Sample Consensus �RANSAC) [23]... |

2215 | A method for registration of 3-d shapes - Besl, McKay - 1992 |

1086 |
Robust regression and outlier detection
- Rousseeuw, Leroy
- 1987
(Show Context)
Citation Context ...C1, is determined using standard methods �see [23], [41], [50]). After minimization, a robust scale value, ^1, is calculated from … ^ A; ^t1† using the median absolute deviation �MAD) scale estimator =-=[42]-=-: 5 ^1 ˆ 1:4826 1 ‡ jP1j 3 median min p2P1 q2C1…p† kq Ap ^ ^t1k 2 r : …12† The factor 1 ‡ 5=…jP1j 3† is an experimentally determined finite-sample correction factor proposed by Rousseeuw and Leroy [42... |

584 |
Robust Statistics. The Approach Based on Influence Functions
- Hampel, R
- 1986
(Show Context)
Citation Context ...ractions well below 50 percent. 4.4 Second Order, Quadratic Model The final level of the hierarchy estimates the quadratic transformation, T2…p; †ˆ X…p†; �see �7) and discussion) using an M-estimator =-=[27]-=-, [50]. We describe first a straightforward instantiation of the M-estimator and then show several important modifications. For each p i 2 P1, let qi ˆ argmin kq q2C1…pi† Api ^ ^t1k 2 : In other words... |

578 | Heirarchical model-based motion estimation - Bergen, Anandan, et al. - 1992 |

567 | Object modeling by registration of multiple range images - Chen, Medioni - 1991 |

473 | A robust technique for matching two uncalibrated images through the recovery of the unknown epipolar geometry
- Zhang, Deriche, et al.
- 1995
(Show Context)
Citation Context ...atch for each transformed p 2 P1 and then taking the median of the resulting squared �error) distances. This differs from other well-known uses of LMSÐe.g., in estimating the fundamental matrix [53], =-=[57]-=-Ðbecause uniqueness in the correspondence set is not �and should not be) enforced at this level of the computation. The objective function in �11) is not differentiable, partially due to use of the me... |

362 | Least median of squares regression - Rousseeuw - 1984 |

350 |
Dynamic 3d models with local and global deformations: deformable superquadrics
- Terzopoulos, Metaxas
- 1991
(Show Context)
Citation Context ...gistration, ranging from low-order models that may include just rotation, translation, and scale to high-order, nonlinear models that include global transformations, and local deformations [6], [51], =-=[52]-=-. Our quadratic model fits somewhere in between, but is more sophisticated than models used in most global motion estimation and mosaic construction algorithms. Similarly, most previous retinal image ... |

322 | Distance transformations in digital images - Borgefors - 1986 |

219 | The Development and Comparison of Robust Methods for Estimating the Fundamental Matrix
- Torr, Murray
- 1997
(Show Context)
Citation Context ... matching and transformation estimation problem in registering retinal images requires a more sophisticated algorithm than has been proposed for related problems such as fundamental matrix estimation =-=[53]-=-, [57]. 4.1 Hierarchical Estimation Algorithm Successfully estimating the parameters of the high-dimensional transformation requires good initialization, especially since correspondences cannot be est... |

219 | Computing occluding and transparent motions - Irani, Rousso, et al. - 1994 |

156 | A New Algorithm for Non-Rigid Point Matching
- Chui, Rangarajan
- 2000
(Show Context)
Citation Context ...that our algorithms do not assume unique correspondences, but instead mix the selection of correct correspondences with robust parameter estimation. Very few algorithms in the literature do this. See =-=[19]-=- for one recent exception. The last step of our estimation algorithm is landmark refinement and correspondence enhancement based on the Fig. 15. Overlapping region of incorrectly registered images whe... |

151 |
Robust regression using iteratively reweighted least-squares
- W, Welsch
- 1977
(Show Context)
Citation Context ...s a scale estimate. Here, we use the Beaton-Tukey biweight function [2] �Fig. 6): …u† ˆ a2 6 1 …1 � u a 2 3 juj a juj >a; a 2 6 where u ˆkq X…p†k=^ is a ªscale normalized residual.º �Typically, a 4:0 =-=[28]-=-.) Equation �14) may be solved using iteratively-reweighted least-squares �IRLS) [28], with weight function w…u† ˆ 0 …u†=u. For the Beaton-Tukey biweight �Fig. 6) w…u† ˆ u 2 2 1 a juj a 0 juj >a: The ... |

130 | Matching 3-D anatomical surfaces with non-rigid deformations using octree-splines
- Szeliski, Lavallee
- 1996
(Show Context)
Citation Context ...for registration, ranging from low-order models that may include just rotation, translation, and scale to high-order, nonlinear models that include global transformations, and local deformations [6], =-=[51]-=-, [52]. Our quadratic model fits somewhere in between, but is more sophisticated than models used in most global motion estimation and mosaic construction algorithms. Similarly, most previous retinal ... |

129 | Robust parameter estimation in computer vision - Stewart - 1999 |

111 | Robust video mosaicing through topology inference and local to global alignment
- Sawhney, Hsu, et al.
- 1998
(Show Context)
Citation Context ...e natural approach is to obtain initial estimates from lower-order models. This suggests a hierarchical estimation strategy, such as adopted in much of the closely related literature [6], [29], [32], =-=[43]-=-, [44]. To estimate , we adopt a three-level hierarchy, beginning with a rough estimate of 2D translations, continuing to an affine estimate, and ending with the quadratic transformation, . At each le... |

110 | True multi-image alignment and its application to mosaicing and lens distortion correction - Sawhney, Kumar - 1999 |

86 | Comparing and evaluating interest points - Schmid, Mohr, et al. - 1998 |

85 | Medical image analysis: progress over two decades and the challenges ahead
- DUNCAN, AYACHE
- 2000
(Show Context)
Citation Context ...em to place the data in a common coordinate system. The data may be twodimensional intensity images [12], range images [10], [18], or volumetric images such as CT scans, MRIs, or confocal stacks [3], =-=[21]-=-. Many techniques have been proposed to solve the registration problem in many different forms. For reviews, see [12], [21], [32]. . A. Can and B. Roysam are with the Department of Electrical, Compute... |

80 | Rapid Automated Tracing and Feature Extraction from Retinal Fundus Images Using Direct Exploratory Algorithms
- Can, Shen, et al.
- 1999
(Show Context)
Citation Context ... of retinal images. The algorithm, which could be generalized to other applications, uses branching points and crossover points in the retinal vasculature detected using a recursive tracing algorithm =-=[14]-=-. The final contribution is a method to refine feature positions while estimating the transformation parameters. This refinement step is necessary for achieving the final accuracy of the registration.... |

79 | A survey of hierarchical non-linear medical image registration - Lester, Arridge - 1999 |

68 | The fitting of power series, meaning polynomials, illustrated on band-spectroscopic data - Beaton, Tukey - 1974 |

52 |
Construction and refinement of panoramic mosaics with global and local alignment
- Shum, Szeliski
- 1998
(Show Context)
Citation Context ...th problems can be addressed as the pairwise registration converges. The transformation estimate can be used as the basis for refining landmark locations and for detecting landmarks previously missed =-=[49]-=-. As already discussed, this can be used to refine the transformation estimate. 5.1 Feature Position Refinement Suppose p i and q j form a landmark feature correspondence and let the estimated transfo... |

42 | Muse: robust surface fitting using unbiased scale estimates - Miller, Stewart - 1996 |

40 |
A multimodal registration algorithm of eye fundus images using vessels detection and hough transform
- Zana, Klein
- 1999
(Show Context)
Citation Context ... average registration error using the affine model is 2.2 pixels and the average registration error using translation only is 5.0 pixels. These results are significant since reported prior work [40], =-=[56]-=- has been based on affine or simpler models, tested on images of size 512 512 or smaller. Matsopoulos et al. [35] studied the use of bilinear and projective transformation models, but could not show i... |

37 |
Distance Transformations
- Borgefors
- 1986
(Show Context)
Citation Context ...each mapped centerline location that falls within the retina image area of Iq, find the distance to the nearest centerline point in Iq. This distance is easily calculated using a digital distance map =-=[11]-=-. The median of these distances is taken as the centerline mapping error. Median statistics are used to discard error measures from missing or spurious centerlines. The other quantity used in evaluati... |

28 | The quadric reference surface: Theory and applications
- Shashua, Toelg
- 1997
(Show Context)
Citation Context ...;y;1† T : Equation �7) has several important properties. It generalizes the 2D affine transformation model induced by the rigid motion of a planar surface [6], which is given by: q ˆ Ap ‡ t: …8† �See =-=[46]-=-, [54] for other recent generalizations.) Observe, also, that �7) is the second-order Taylor series expansion of the general image transformation equation, while the affine model �8) is the first-orde... |

26 |
Automatic retinal image registration scheme using global optimization techniques
- Matsopoulos, Mouravliansky, et al.
- 1999
(Show Context)
Citation Context ...surgery and other related procedures [5], [8], [9]. Despite the importance of the problem and prior research effort, retinal image registration has remained a difficult problem [5], [20], [26], [33], =-=[35]-=-. Several challenges must be addressed in developing reliable automatic registration techniques. These challenges are special cases of difficulties that arise in other applications. . The surface of t... |

25 | Robust Statistics: The Approach Based on In Functions - Hampel, Ronchetti, et al. - 1986 |

25 | Registration of stereo and temporal images of the retina - Ritter, Owens, et al. - 1999 |

24 | Optimal scheduling of tracing computations for real-time vascular landmark extraction from retinal fundus images
- Shen, Roysam, et al.
(Show Context)
Citation Context ...n of the algorithm has been implemented to run at frame rates on an SGI workstation [1]. In a recent paper, we have shown how to optimally prioritize the exploratory tracing for real-time performance =-=[47]-=-. 4 ESTIMATING THE TRANSFORMATION PARAMETERS AND THE FEATURE CORRESPONDENCES The next step is to use the vascular landmarks to estimate the transformation parameters for a pair of retinal images.sCAN ... |

16 | Robust hierarchical algorithm for constructing a mosaic from images of the curved human retina - Can, Stewart, et al. - 1999 |

16 | A statistical framework for long-range feature matching in uncalibrated image mosaicing
- Cham, Cipolla
- 1998
(Show Context)
Citation Context ...els and associated robust estimation techniques. A number of other motion and registration algorithms in the computer vision and medical image analysis literatures have used a similar model hierarchy =-=[17]-=-, [29], [32], [43], [44], [51]; doing so is a natural approach. Of particular interest here, Cham and Cipolla [17] have developed a hierarchical, Bayesian method combining Kalman filters and RANSAC me... |

15 | Image processing algorithms for retinal montage synthesis, mapping, and real-time location determination
- Becker, Can, et al.
- 1998
(Show Context)
Citation Context ...nsity structure is relatively homogeneous. Third, the vasculature and its bifurcations are currently used as spatial landmarks by ophthalmologists and in much of the published literature in this area =-=[4]-=-, [35], [56]. Finally, except for detachments, the retina is rigidly attached to the back of the eye. Hence, the branching and crossover points of the blood vessel structure serve as our features. The... |

15 | A feature-based algorithm for joint, linear estimation of high-order image-to-mosaic transformations: mosaicing the curved human retina - Can, Stewart, et al. - 2002 |

14 |
A Survey of Image Registration
- Brown
- 1992
(Show Context)
Citation Context ...the registration problem requires estimating the transformation�s) between data sets and applying them to place the data in a common coordinate system. The data may be twodimensional intensity images =-=[12]-=-, range images [10], [18], or volumetric images such as CT scans, MRIs, or confocal stacks [3], [21]. Many techniques have been proposed to solve the registration problem in many different forms. For ... |

14 | A feature-based technique for joint linear estimation of high-order image-to-mosaic transformations: mosaicing the curved human retina
- Can, Stewart, et al.
- 2002
(Show Context)
Citation Context ...c formation. This paper is the first of two devoted to solving the retina image registration problem. The current paper addresses the problem of registering a pair of retinal images. The second paper =-=[16]-=- addresses the problem of simultaneously registering multiple retinal images to form a mosaic. Together, they describe a set of algorithms that are currently being tailored to several ophthalmic appli... |

12 | Q-warping: Direct computation of quadratic reference surfaces - Shashua, Wexler - 2001 |

11 | Real-time algorithm for retinal tracking - Markow, Rylander, et al. - 1993 |

9 | Digital image processing for ocular fundus images - Goldbaum, Katz, et al. - 1990 |

9 | Automated registration of digital ocular fundus images for comparison of lesions - Goldbaum, Kouznetsova, et al. - 1993 |

8 |
ªA Method for Registration of 3-D Shapes,º
- Besl, McKay
- 1992
(Show Context)
Citation Context ...oblem requires estimating the transformation�s) between data sets and applying them to place the data in a common coordinate system. The data may be twodimensional intensity images [12], range images =-=[10]-=-, [18], or volumetric images such as CT scans, MRIs, or confocal stacks [3], [21]. Many techniques have been proposed to solve the registration problem in many different forms. For reviews, see [12], ... |

8 | Constructing retinal fundus photomontages. A new computer-based method,” Investigat - Mahurkar, Vivino, et al. - 1996 |

7 |
ªHierarchical Model-Based Motion Estimation,º
- Bergen, Anandan, et al.
- 1992
(Show Context)
Citation Context ....; i6†, i ˆ 1; 2, and X…p† ˆ…x 2 ; xy; y 2 ;x;y;1† T : Equation �7) has several important properties. It generalizes the 2D affine transformation model induced by the rigid motion of a planar surface =-=[6]-=-, which is given by: q ˆ Ap ‡ t: …8† �See [46], [54] for other recent generalizations.) Observe, also, that �7) is the second-order Taylor series expansion of the general image transformation equation... |

7 | Projective registration with dierence decomposition - Gleicher - 1997 |

6 |
J: Computer-assisted, interactive fundus image processing for macular drusen quantitation. Ophthalmology 106
- Shin, Javornik, et al.
- 1998
(Show Context)
Citation Context ...scars and burns [7]. Two images taken of the same eye but at different times can be registered to track the progress of diseases such as macular degeneration, glaucoma, and AIDS/CMV retinopathy [31], =-=[48]-=-. A series of images of the same retina can be registered to form a mosaic image, giving a complete view of the retina [33]. Finally, real-time �frame-rate) registration methods may be used as the bas... |

6 | Registration of ocular fundus images - Cideciyan - 1995 |

5 |
Automated 3-D montage synthesis from laser-scanning confocal images: application to quantitative tissue-level cytological analysis
- Becker, Ancin, et al.
- 1996
(Show Context)
Citation Context ...ng them to place the data in a common coordinate system. The data may be twodimensional intensity images [12], range images [10], [18], or volumetric images such as CT scans, MRIs, or confocal stacks =-=[3]-=-, [21]. Many techniques have been proposed to solve the registration problem in many different forms. For reviews, see [12], [21], [32]. . A. Can and B. Roysam are with the Department of Electrical, C... |

5 |
A Survey of Hierarchical Non-Linear
- Lester, Arridge
- 1999
(Show Context)
Citation Context ..., or volumetric images such as CT scans, MRIs, or confocal stacks [3], [21]. Many techniques have been proposed to solve the registration problem in many different forms. For reviews, see [12], [21], =-=[32]-=-. . A. Can and B. Roysam are with the Department of Electrical, Computer, and Systems Engineering, Rensselaer Polytechnic Institute, Troy, NY 12180-3590. E-mail: roysab@rpi.edu, alican@alum.rpi.edu. .... |