## Near-optimal detection of geometric objects by fast multiscale methods (2005)

### Cached

### Download Links

- [www.isye.gatech.edu]
- [www2.isye.gatech.edu]
- [www2.isye.gatech.edu]
- [www.isye.gatech.edu]
- [www.isye.gatech.edu]
- [www2.isye.gatech.edu]
- [www2.isye.gatech.edu]
- DBLP

### Other Repositories/Bibliography

Venue: | IEEE Trans. Inform. Theory |

Citations: | 23 - 7 self |

### BibTeX

@ARTICLE{Arias-castro05near-optimaldetection,

author = {Ery Arias-castro and David L. Donoho and Xiaoming Huo and Senior Member},

title = {Near-optimal detection of geometric objects by fast multiscale methods},

journal = {IEEE Trans. Inform. Theory},

year = {2005},

volume = {51}

}

### OpenURL

### Abstract

Abstract—We construct detectors for “geometric ” objects in noisy data. Examples include a detector for presence of a line segment of unknown length, position, and orientation in two-dimensional image data with additive white Gaussian noise. We focus on the following two issues. i) The optimal detection threshold—i.e., the signal strength below which no method of detection can be successful for large dataset size. ii) The optimal computational complexity of a near-optimal detector, i.e., the complexity required to detect signals slightly exceeding the detection threshold. We describe a general approach to such problems which covers several classes of geometrically defined signals; for example, with one-dimensional data, signals having elevated mean on an interval, and, in-dimensional data, signals with elevated mean on a rectangle, a ball, or an ellipsoid. In all these problems, we show that a naive or straightforward approach leads to detector thresholds and algorithms which are asymptotically far away from optimal. At the same time, a multiscale geometric analysis of these classes of objects allows us to derive asymptotically optimal detection thresholds and fast algorithms for near-optimal detectors. Index Terms—Beamlets, detecting hot spots, detecting line segments, Hough transform, image processing, maxima of Gaussian processes, multiscale geometric analysis, Radon transform. I.

### Citations

3719 |
Stochastic Relaxation, Gibbs Distributions and the Bayesian Restoration of Images
- Geman, Geman
- 1984
(Show Context)
Citation Context ...ly [7], [35], [38], [66], [69]. At the same time, there is a very extensive literature on statistical image reconstruction from noisy data. Since the mid-1980s, work of Besag [11] and Geman and Geman =-=[39]-=-, Bayesian methods have been widely used in reconstruction of images from noisy data. Based on a model for the object(s) to be estimated/identified in the image, together with a noise model, Bayesian ... |

922 | On the statistical analysis of dirty pictures
- Besag
- 1986
(Show Context)
Citation Context ... multiscale and run rapidly [7], [35], [38], [66], [69]. At the same time, there is a very extensive literature on statistical image reconstruction from noisy data. Since the mid-1980s, work of Besag =-=[11]-=- and Geman and Geman [39], Bayesian methods have been widely used in reconstruction of images from noisy data. Based on a model for the object(s) to be estimated/identified in the image, together with... |

849 | E.: “Theory of edge detection
- Marr, Hildreth
- 1980
(Show Context)
Citation Context ...pic areas over the years, only a tiny fraction of which can be cited here. Some of that literature even uses cognate language: “detectors” for edges go back to classical work of the late 1970s, e.g., =-=[55]-=-; more recent literature is even said to develop “object detectors” which are multiscale and run rapidly [7], [35], [38], [66], [69]. At the same time, there is a very extensive literature on statisti... |

804 | Robust real-time object detection
- Viola, Jones
- 2001
(Show Context)
Citation Context ...ctors” for edges go back to classical work of the late 1970s, e.g., [55]; more recent literature is even said to develop “object detectors” which are multiscale and run rapidly [7], [35], [38], [66], =-=[69]-=-. At the same time, there is a very extensive literature on statistical image reconstruction from noisy data. Since the mid-1980s, work of Besag [11] and Geman and Geman [39], Bayesian methods have be... |

583 | Active Contours - Blake, Isard - 1998 |

567 |
Convergence of Stochastic Processes
- Pollard
- 1984
(Show Context)
Citation Context ... Lemma 2.8: . (2.9) This, together with (2.6), will allow us to show that and behave similarly. To prove (2.9), we need the following well-known fact about extreme values of Gaussian processes ([52], =-=[57]-=-, [62]). Lemma 2.9: Let be (possibly dependent) variables with all . Then Applying this to , set As , we get from (2.10) that for which yields (2.9). (2.10)sARIAS-CASTRO et al.: NEAR-OPTIMAL DETECTION... |

382 |
Empirical Processes with Applications to Statistics
- Shorack, Wellner
- 1986
(Show Context)
Citation Context ...(2.9) This, together with (2.6), will allow us to show that Z ∗ [In] and Z ∗ [Jn,ℓ] behave similarly. To prove (2.9), we need the following well-known fact about extreme values of Gaussian processes (=-=[65, 51]-=-): Lemma 2.9 Let w1, ..., wm be (possibly dependent) N (0, σ 2 i ) variables with all σi ≤ σ. Then, Applying this to ∆n,ℓ, set P (max{w1, ..., wm} > � 2 log(m) · σ) ≤ 1/ � 4π log(m). (2.10) m = #{I − ... |

280 |
Deformable templates using large deformation kinematics,” Image Processing
- Christensen, Rabbitt, et al.
- 1996
(Show Context)
Citation Context ... with a noise model, Bayesian methods rely on a posterior mode or posterior expectations to find a reconstructed object, often computed by Monte Carlo techniques [12], [39]. Deformable templates [8], =-=[21]-=-, [34], [44], [45], [56], [58]–[60] and marked point processes [64], [68] are often used as models for deriving Bayesian methods. 2) Differences in Goals: The prior literature in these two fields, whi... |

246 | Minimax estimation via wavelet shrinkage
- Donoho, Johnstone
- 1998
(Show Context)
Citation Context ..., [48]. For and , hypotheses and merge asymptotically, meaning that every sequence of tests has Type I Type II errors tending to . See also related work of Ingster [49], [50], of Donoho and Johnstone =-=[31]-=-, and of Jin [30], which elaborates this fact in various ways. Now the hypothesis is a subhypothesis of the laid out in the statement of Theorem 2.3, so the problem of distinguishing from must be at l... |

230 |
Statistical Estimation: Asymptotic Theory
- Ibragimov, Has’minskii
- 1981
(Show Context)
Citation Context ...hite noise might have a single nonzero mean in an unknown position may be called the “needle in a haystack” problem, and it has been known, according to [17], since work of Ibragimov and Khas’minskii =-=[47]-=-, [48]. For and , hypotheses and merge asymptotically, meaning that every sequence of tests has Type I Type II errors tending to . See also related work of Ingster [49], [50], of Donoho and Johnstone ... |

215 |
General Pattern Theory
- Grenander
- 1993
(Show Context)
Citation Context ...ogether with a noise model, Bayesian methods rely on a posterior mode or posterior expectations to find a reconstructed object, often computed by Monte-Carlo techniques [36, 10]. Deformable templates =-=[58, 43, 44, 61, 62, 63]-=- and marked point processes [67, 53] are often used as models for deriving Bayesian methods. 1.4.2 Differences in Goals The prior literature in these two fields, while fundamental and very useful, see... |

201 |
Representations of knowledge in complex systems (with discussion
- Grenander, Miller
- 1994
(Show Context)
Citation Context ...l, Bayesian methods rely on a posterior mode or posterior expectations to find a reconstructed object, often computed by Monte Carlo techniques [12], [39]. Deformable templates [8], [21], [34], [44], =-=[45]-=-, [56], [58]–[60] and marked point processes [64], [68] are often used as models for deriving Bayesian methods. 2) Differences in Goals: The prior literature in these two fields, while fundamental and... |

176 | On edge detection
- Torre, Poggio
- 1986
(Show Context)
Citation Context ... “detectors” for edges go back to classical work of the late 1970s, e.g., [55]; more recent literature is even said to develop “object detectors” which are multiscale and run rapidly [7], [35], [38], =-=[66]-=-, [69]. At the same time, there is a very extensive literature on statistical image reconstruction from noisy data. Since the mid-1980s, work of Besag [11] and Geman and Geman [39], Bayesian methods h... |

134 |
Bayesian computation and stochastic systems (with discussion
- Besag, Green, et al.
- 1995
(Show Context)
Citation Context ...ated/identified in the image, together with a noise model, Bayesian methods rely on a posterior mode or posterior expectations to find a reconstructed object, often computed by Monte Carlo techniques =-=[12]-=-, [39]. Deformable templates [8], [21], [34], [44], [45], [56], [58]–[60] and marked point processes [64], [68] are often used as models for deriving Bayesian methods. 2) Differences in Goals: The pri... |

126 |
Wedgelets: Nearly minimax estimation of edges
- Donoho
- 1999
(Show Context)
Citation Context ...n , and let . is a dyadically organized set of line segments, occupying a range of orientations, scales, and locations. This system has been discussed at length by Donoho and Huo [28], [29]; see also =-=[27]-=-sARIAS-CASTRO et al.: NEAR-OPTIMAL DETECTION OF GEOMETRIC OBJECTS BY FAST MULTISCALE METHODS 2415 Fig. 3. Dyadic squares marked with vertices, and several beamlets. Fig. 4. Approximating a line segmen... |

125 |
Structural image restoration through deformable template
- Amit, Grenander, et al.
- 1991
(Show Context)
Citation Context ...ether with a noise model, Bayesian methods rely on a posterior mode or posterior expectations to find a reconstructed object, often computed by Monte Carlo techniques [12], [39]. Deformable templates =-=[8]-=-, [21], [34], [44], [45], [56], [58]–[60] and marked point processes [64], [68] are often used as models for deriving Bayesian methods. 2) Differences in Goals: The prior literature in these two field... |

121 | R.: Fast multiscale image segmentation
- Sharon, Brandt, et al.
- 2000
(Show Context)
Citation Context ...67], and so on. There is an extensive cognate literature on line detection in computer vision—extending back decades, for example, under the label “Hough transform”—for which citations in [26], [42], =-=[61]-=- may provide helpful pointers and applications references. This paper derives asymptotic properties of: i) the signal amplitude at which detection in noisy data is possible; and ii) the computational ... |

117 |
Extremes and related properties of random sequences and processes. Springer Series in Statistics
- Leadbetter, Lindgren, et al.
- 1983
(Show Context)
Citation Context ...(2.9) This, together with (2.6), will allow us to show that Z ∗ [In] and Z ∗ [Jn,ℓ] behave similarly. To prove (2.9), we need the following well-known fact about extreme values of Gaussian processes (=-=[65, 51]-=-): Lemma 2.9 Let w1, ..., wm be (possibly dependent) N (0, σ 2 i ) variables with all σi ≤ σ. Then, Applying this to ∆n,ℓ, set P (max{w1, ..., wm} > � 2 log(m) · σ) ≤ 1/ � 4π log(m). (2.10) m = #{I − ... |

97 | Coarse-to-fine face detection
- Fleuret, Geman
(Show Context)
Citation Context ...te language: “detectors” for edges go back to classical work of the late 1970s, e.g., [55]; more recent literature is even said to develop “object detectors” which are multiscale and run rapidly [7], =-=[35]-=-, [38], [66], [69]. At the same time, there is a very extensive literature on statistical image reconstruction from noisy data. Since the mid-1980s, work of Besag [11] and Geman and Geman [39], Bayesi... |

85 | Higher criticism for detecting sparse heterogeneous mixtures
- Donoho, Jin
(Show Context)
Citation Context ... hypotheses and merge asymptotically, meaning that every sequence of tests has Type I Type II errors tending to . See also related work of Ingster [49], [50], of Donoho and Johnstone [31], and of Jin =-=[30]-=-, which elaborates this fact in various ways. Now the hypothesis is a subhypothesis of the laid out in the statement of Theorem 2.3, so the problem of distinguishing from must be at least as hard. The... |

78 | Representation and detection of deformable shapes
- Felzenszwalb
- 2003
(Show Context)
Citation Context ...a noise model, Bayesian methods rely on a posterior mode or posterior expectations to find a reconstructed object, often computed by Monte Carlo techniques [12], [39]. Deformable templates [8], [21], =-=[34]-=-, [44], [45], [56], [58]–[60] and marked point processes [64], [68] are often used as models for deriving Bayesian methods. 2) Differences in Goals: The prior literature in these two fields, while fun... |

62 |
Bump hunting in high dimensional data
- Friedman, Fisher
- 1999
(Show Context)
Citation Context ...mplitude. Letting denote the set of -dimensional rectangles, we are interested in testing versus , . This is related to problems of finding “high activity regions” [6], [46], and multivariate “bumps” =-=[37]-=-. If a specific rectangle is the only one under consideration, we would be interested in the simple hypothesis-testing problem versus , where the particular is specified. The Neyman–Pearson test could... |

62 | Detection of Linear Features in SAR Images: Application to Road Network Extraction
- Tupin, Maitre, et al.
- 1998
(Show Context)
Citation Context ...tation, or location of the “V” [22]. We envision that detecting line segments is a component task in more ambitious settings: detecting cracks in materials [54], identifying streams and roadbeds [5], =-=[67]-=-, and so on. There is an extensive cognate literature on line detection in computer vision—extending back decades, for example, under the label “Hough transform”—for which citations in [26], [42], [61... |

60 | Rectifiable sets and the traveling salesman problem - Jones - 1990 |

55 | Beamlets and multiscale image analysis
- Donoho, Huo
- 2002
(Show Context)
Citation Context ...ne segments in the square is O(n 2 log(n)), not O(n 4 ). To understand this, we have developed tools that view the space of line segments from a multiscale viewpoint. Our approach has four components =-=[27]-=-: 4s1. Beamlets. A dictionary of multiscale line segments at a range of orientations, and at dyadically-organized scales and locations. It has cardinality O(n 2 log(n)). 2. Beamlet chaining. Arbitrary... |

49 | A coarse-to-fine strategy for multiclass shape detection
- Amit, Geman, et al.
(Show Context)
Citation Context ...cognate language: “detectors” for edges go back to classical work of the late 1970s, e.g., [55]; more recent literature is even said to develop “object detectors” which are multiscale and run rapidly =-=[7]-=-, [35], [38], [66], [69]. At the same time, there is a very extensive literature on statistical image reconstruction from noisy data. Since the mid-1980s, work of Besag [11] and Geman and Geman [39], ... |

49 | Fast slant stack: A notion of radon transform for data in a cartesian grid which is rapidly computible, algebraically exact, geometrically faithful and invertible
- Averbuch, Coifman, et al.
- 2001
(Show Context)
Citation Context ... of approximate recursive computation in beamlet-like systems has been proposed by Brandt and Dym [16] and by Götze and Druckenmiller [42]. Other effective approaches include fast Radon transforms in =-=[10]-=-, [14]. If , then any such fast algorithm does not give perfect identity, but instead high correlation where as . The point is that for sufficiently close to , an acceptable approximation to can be co... |

49 | Analysis of and on uniformly rectifiable sets - David, Semmes - 1994 |

45 | Meaningful alignments
- Desolneux, Moisan, et al.
(Show Context)
Citation Context ...dbeds [5], [67], and so on. There is an extensive cognate literature on line detection in computer vision—extending back decades, for example, under the label “Hough transform”—for which citations in =-=[26]-=-, [42], [61] may provide helpful pointers and applications references. This paper derives asymptotic properties of: i) the signal amplitude at which detection in noisy data is possible; and ii) the co... |

40 | Multiscale Scientific Computation: Review 2001”. In
- Brandt
- 2002
(Show Context)
Citation Context ...s for Bayesian inference seems to prevent any near-optimality from the computational complexity viewpoint. Perhaps a multiscale geometric analysis approach could speed up such methods, as Achi Brandt =-=[14]-=- has often argued for methods inspired by statistical physics. Moreover, the detection thresholds and other theoretical properties implied by such methods seem not to have been studied so far. 1.4.4 I... |

39 |
Scan Statistics
- Glaz, Naus, et al.
- 2001
(Show Context)
Citation Context ...N INFORMATION THEORY, VOL. 51, NO. 7, JULY 2005 within a specified proportion of errors. More generally, the scan statistic is a widely used tool in statistical sequence analysis; see the books [40], =-=[41]-=-. It measures the maximal number of occurrences in any interval of a given length. The statistic we are using is a multiscale scan statistic, which considers all possible lengths and positions and com... |

39 | pattern theory: A Mathematical Study of Regular Structures - General - 1993 |

32 |
Single-particle imaging of macromolecules by cryo-electron microscopy
- Frank
(Show Context)
Citation Context ... of image-processing problems, including target recognition [18] and medical imaging. Detecting disks and tilted rectangles in very noisy data is currently a priority in cryo-electron microscopy [3], =-=[36]-=-, [65], [70]. Our theory places fundamental limits on attainable performance, and provides a standard of comparison for practical work. A. Line Segment Detection Suppose we have an -by- array of pixel... |

29 | The Erdös-Rényi strong law for pattern matching with a given proportion of mismatches. The Annals of Probability 17
- Arratia, Waterman
- 1989
(Show Context)
Citation Context ...at works for much more general classes of sets than intervals. Farther afield, but sure to come up in a close examination of our topic, is work in genome sequence analysis. Waterman and collaborators =-=[9]-=-, [23], [24] studied asymptotic behavior of detectors searching for squares and/or rectangles of specified values in discrete-valued data. They define matching not in terms of statistical significance... |

26 |
Detection filters and algorithm fusion for ATR
- Casasent, Ye
- 1997
(Show Context)
Citation Context ...efficient algorithms for near-optimal detection in such settings. In practice, detecting geometric objects is an important task in a variety of image-processing problems, including target recognition =-=[18]-=- and medical imaging. Detecting disks and tilted rectangles in very noisy data is currently a priority in cryo-electron microscopy [3], [36], [65], [70]. Our theory places fundamental limits on attain... |

23 |
Differential Snakes for Change Detection in Road Segments." Photogrammetric Engineering & Remote Sensing 67.12
- Agouris, Stefanidis, et al.
- 2001
(Show Context)
Citation Context ...orientation, or location of the “V” [22]. We envision that detecting line segments is a component task in more ambitious settings: detecting cracks in materials [54], identifying streams and roadbeds =-=[5]-=-, [67], and so on. There is an extensive cognate literature on line detection in computer vision—extending back decades, for example, under the label “Hough transform”—for which citations in [26], [42... |

21 |
A fast discrete approximation algorithm for the Radon transform
- Brady
- 1998
(Show Context)
Citation Context ...proximate recursive computation in beamlet-like systems has been proposed by Brandt and Dym [16] and by Götze and Druckenmiller [42]. Other effective approaches include fast Radon transforms in [10], =-=[14]-=-. If , then any such fast algorithm does not give perfect identity, but instead high correlation where as . The point is that for sufficiently close to , an acceptable approximation to can be computed... |

18 | Beamlet pyramids: a new form of multiresolution analysis, suited for extracting lines, curves and objects from very noisy image data
- Donoho, Huo
- 2000
(Show Context)
Citation Context ... of marked vertices in , and let . is a dyadically organized set of line segments, occupying a range of orientations, scales, and locations. This system has been discussed at length by Donoho and Huo =-=[28]-=-, [29]; see also [27]sARIAS-CASTRO et al.: NEAR-OPTIMAL DETECTION OF GEOMETRIC OBJECTS BY FAST MULTISCALE METHODS 2415 Fig. 3. Dyadic squares marked with vertices, and several beamlets. Fig. 4. Approx... |

16 | Bayesian object recognition with Baddeley’s delta loss
- Rue, Syversveen
- 1998
(Show Context)
Citation Context ...ds rely on a posterior mode or posterior expectations to find a reconstructed object, often computed by Monte Carlo techniques [12], [39]. Deformable templates [8], [21], [34], [44], [45], [56], [58]–=-=[60]-=- and marked point processes [64], [68] are often used as models for deriving Bayesian methods. 2) Differences in Goals: The prior literature in these two fields, while fundamental and very useful, see... |

16 |
A gibbs point process for road extraction in remotely sensed images
- Stoica, Descombes, et al.
- 2004
(Show Context)
Citation Context ...osterior expectations to find a reconstructed object, often computed by Monte Carlo techniques [12], [39]. Deformable templates [8], [21], [34], [44], [45], [56], [58]–[60] and marked point processes =-=[64]-=-, [68] are often used as models for deriving Bayesian methods. 2) Differences in Goals: The prior literature in these two fields, while fundamental and very useful, seems to pursue different goals tha... |

16 |
A fast digital radon transform- an efficient means for evaluating the hough transform
- Gotz, Druckmuller
- 1995
(Show Context)
Citation Context ...oadbeds [68], and so on. There is an extensive cognate literature on line detection in computer vision – extending back decades, for example under the label ‘Hough Transform’ – for which citations in =-=[64, 25, 41]-=- may provide helpful pointers and applications references. This paper derives asymptotic (n → ∞) properties of: (i) the signal amplitude at which detection in noisy data is possible; and (ii) the comp... |

15 |
Using the generalized likelihood ratio statistic for sequential detection of a change-point
- Siegmund, Venkatraman
- 1995
(Show Context)
Citation Context ... sequential analysis, there has been interest for some time in probability distributions for quantities of the form where is a white Gaussian noise. See, for example, work of Siegmund and Venkatraman =-=[63]-=- and Lai and Chan [19], [20]. This problem is closely related to our Section II, which involves detecting a one-dimensional signal with an elevated mean. The techniques developed in sequential analysi... |

15 | Differential Snakes for Change Detection - Agouris, Stefanidis - 2001 |

14 |
Fast calculation of multiple line integrals
- Brandt, Dym
- 1999
(Show Context)
Citation Context ...T as it is for constructing fast algorithms to approximate it. Multiscale approaches to detect linear features in two dimensions have been invoked for some time. Early examples include work of Brandt =-=[16]-=-, [61], Horn and collaborators [4], [33]. Later we will discuss related literature and our relationship to it. C. A Range of Geometric Problems While our exposition began with the problem of detecting... |

13 |
Some problems of nonparametric estimation in Gaussian white noise. Teor. Veroyatnost. i Primenen
- Ibragimov, Nemirovskiĭ, et al.
- 1986
(Show Context)
Citation Context ...oise might have a single nonzero mean in an unknown position may be called the “needle in a haystack” problem, and it has been known, according to [17], since work of Ibragimov and Khas’minskii [47], =-=[48]-=-. For and , hypotheses and merge asymptotically, meaning that every sequence of tests has Type I Type II errors tending to . See also related work of Ingster [49], [50], of Donoho and Johnstone [31], ... |

13 |
On some problems of hypothesis testing leading to infinitely divisible distributions
- Ingster
- 1997
(Show Context)
Citation Context ...k of Ibragimov and Khas’minskii [47], [48]. For and , hypotheses and merge asymptotically, meaning that every sequence of tests has Type I Type II errors tending to . See also related work of Ingster =-=[49]-=-, [50], of Donoho and Johnstone [31], and of Jin [30], which elaborates this fact in various ways. Now the hypothesis is a subhypothesis of the laid out in the statement of Theorem 2.3, so the problem... |

12 | Fast x-ray and beamlet transforms for three-dimensional data - Donoho, Levi - 2004 |

12 | Bayesian object identification
- Rue, Hurn
- 1999
(Show Context)
Citation Context ...methods rely on a posterior mode or posterior expectations to find a reconstructed object, often computed by Monte Carlo techniques [12], [39]. Deformable templates [8], [21], [34], [44], [45], [56], =-=[58]-=-–[60] and marked point processes [64], [68] are often used as models for deriving Bayesian methods. 2) Differences in Goals: The prior literature in these two fields, while fundamental and very useful... |

11 |
Localized radon transform-based detection of ship wakes in sarimages. Geoscience and Remote Sensing
- Copeland, Ravichandran, et al.
- 1995
(Show Context)
Citation Context ...ects that the imagery might sometimes contain a very faint “V” shape against a highly cluttered background; there is no real reason to expect any particular scale, orientation, or location of the “V” =-=[22]-=-. We envision that detecting line segments is a component task in more ambitious settings: detecting cracks in materials [54], identifying streams and roadbeds [5], [67], and so on. There is an extens... |

11 | Boundary detection through dynamic polygons
- Pievatolo, Green
- 1998
(Show Context)
Citation Context ...esian methods rely on a posterior mode or posterior expectations to find a reconstructed object, often computed by Monte Carlo techniques [12], [39]. Deformable templates [8], [21], [34], [44], [45], =-=[56]-=-, [58]–[60] and marked point processes [64], [68] are often used as models for deriving Bayesian methods. 2) Differences in Goals: The prior literature in these two fields, while fundamental and very ... |