W. T. Freeman. Steerable Filters and Local Analysis of Image Structure. PhD thesis, Massachusetts Institute of Technology, Cambridge, MA, 1992. URL: http: // citeseer. nj. nec. com/ freeman92steerable. html  Home/Search   Document Details and Download   Summary   Related Articles   Check

....2.3 Local Orientation and Edge Quality The response modulus in the six channels is used to determine the local orientation of the underlying structure. We use the real and imaginary coefficients of the first Fourier harmonic to approximate the local orientation, similar to the approaches of [4, 10]) loc = 1 tan Gamma1 Im Re ; where loc 2 [ Gamma ; Because we use filters that are polar separable in the Fourier domain, the response magnitudes in the different orientation channels are entirely determined by the orientation of an edge line and the orientation tuning of ....

W.T. Freeman. Steerable Filters and Local Analysis of Image Structure. PhD thesis, MIT, Media Laboratory, Cambridge MA, 1992.

....Gabel and Roberts [25] ffl Section 3.2.1 The z transform, Stark [80] ffl Section 3.2.4 Prony s method, Lanczos [43] ffl Section 3.2. 4 The non linear least squares problem, Lim [44] xi ffl Since the filter design and implementation presented in Chapter 3 is based on the work of Freeman [23, 24] and Deriche [16] selected formulae and workings from these papers are reproduced here. Original material presented in this thesis is the work of the author. xii Acknowledgments I would like to thank my supervisor Robyn Owens for her support and patience throughout my long tenure as her ....

....of the energy model to features in multiple dimensions. Papers by Robbins [73] Robbins and Owens [71] Kovesi [40] Pudney et al. 66] Perona [61] and Heitger et al. 31] all concentrate on applying the one dimensional results at a number of orientations. Canny [8] Deriche [14] and Freeman [24, 23] have also applied multi orientation analysis to edge detection, a technique that was suggested much earlier by Marr and Hildreth [47] The following three subsections (2.2.2 to 2.2.4) discuss the development of the phase CHAPTER 2. LOCAL ENERGY FEATURE DETECTION 18 congruency model of feature ....

[Article contains additional citation context not shown here]

W. T. Freeman. Steerable Filters and Local Analysis of Image Structure. PhD thesis, Architecture and Planning, MIT, June 1992.

.... that the human visual system exploits a range of regularities in image structure to solve this task [12, 14, 28] Inspired in part by these studies, algorithms have been developed to apply continuity, cocircularity and smoothness constraints to organize local edges into extended contours [2, 10,17,19,22,24]. However, despite persuasive psychophysical demonstrations of the role of contour closure in perceptual organization [6, 7, 14, 15, 25] closure constraints for computer vision algorithms have been largely ignored (although see [1, 11] This is surprising, since existing algorithms, while ....

W.T. Freeman. Steerable Filters and Local Analysis of Image Structure. PhD thesis, MIT Media Lab, 1992.

....features are more general in their definition 1 Note that in real images, the occluding surface could be on either side of the continuing line depending on whether the terminating line corresponds to a surface boundary or a texture edge in the scene. For more details, see Freeman s PhD thesis [15]. CHAPTER 2. REVIEW OF 2D FEATURE DETECTION 7 (B) FORK or Y (C) ARROW (D) T (F) PEAK (H) MULTI (I) K (A) L (G) X (E) MATCHED T Figure 2: The nine classes of vertices enumerated by Guzman with the links between regions belonging to the same body. CHAPTER 2. REVIEW OF 2D FEATURE ....

....than any other 2D feature detector that the author has been able to investigate. An example of the quality of the feature detection and localization provided by his scheme is shown in Figure 7. Freeman also presented an approach to 2D feature detection based on oriented energy filters [15]. Like Heitger, he also used derivatives of energy at different orientations to extract 2D image features. The goal of his work was to develop a system to form an initial interpretation of the physical origin of observed image structures . The emphasis of his work therefore was on the ....

[Article contains additional citation context not shown here]

W. T. Freeman. Steerable Filters and Local Analysis of Image Structure. PhD thesis, MIT Media Lab. TR-190, June 1992.

....extended the Saliency Network. Shashua and Ullman [26] developed a method for grouping which, based on the same computation, groups together curve pieces that mutually prefer each other over other candidates. Their new method, however, is not guaranteed to converge to the optimal solution. Freeman [7] used the Saliency Network to detect salient curves and extract junctions in the output of steerable filters. He reported that the method had serious problems with discretization and grouping, similar to some of the problems we describe in Sections 5 and 6. In addition, Subirana Vilanova and Sung ....

....9: Discretizing a circle with a regular polygon. Consequently, we obtain the paradoxical result that the corner is more salient than the nearly straight continuation. Hence straight lines oriented such that they deviate slightly from horizontal will often be less salient than corners. Freeman [7] made a similar observation concerning this aliasing issue when he considered applying the Saliency Network to grouping. The discretization problem is carried over to other, more complicated examples. Consider a circle of radius r. When r is sufficiently small, the circle can be approximated by a ....

Freeman, W. T., "Steerable Filters and Local Analysis of Image Structure," Ph. D. Thesis, M. I. T. Media Lab, Technical Report #190, June, 1992.

....the image using a combination of the magnitudes of the outputs of a set of directional filters convolved with the image in the spatial domain. Unlike the works mentioned above, the implementation used here and in [12] extracts orientation information over multiple scales using a steerable pyramid [14], then combines the orientations from different scales and decides which are dominant perceptually, as determined by a human study. The bottom level of the steerable pyramid (level0) is the original image, and each higher level is obtained by filtering and subsampling the previous level. At each ....

W. T. Freeman. Steerable Filters and Local Analysis of Image Structure. PhD thesis, Media Arts and Sciences, MIT, 1992.

....on the original image. Locations with added white are those where a T junction has been found. Although the primitive Tjunction detector gives reasonable answers for this particular image, it is by no means a general purpose junction finder. More sophisticated junction finders are described in (Freeman, 1992). In the locations in which T junctions were found we set (x; y) 1, i.e. these locations are definitely outliers. The results shown in figure 3 43 are obtained with the same parameters oe N ; oe P as in figure 3 38 but now the solution that maximizes the posterior probability is the one ....

....the image in 3 37b (the output of the detector is superimposed on the original image) We used a very rudimentary t junction detector here that searches for regions with a tri modal intensity distribution. More sophisticated and biologically plausible detectors have been proposed elsewhere (e.g. (Freeman, 1992)) Our goal here is to show how the outputs of such a detector could influence the motion grouping. In our formulation terminators are weighted in accordance with their probability of being extrinsic. That is, the formulation allows a gradual distinction between intrinsic and extrinsic features, ....

Freeman, W. T. (1992). Steerable Filters and Local Analysis of Image Structure.

....Additional observations may support this decision. With regard to computation time, edge detection methods are intrinsically parallel processes. iv Moreover, design of local filters for edge detection features a strong theoretical foundation in signal processing theory and statistical analysis [29] [32] Finally, neurophysiology may provide useful information on how the mammalian Primary Visual Cortex (PVC) employs multi scale filter banks in low level vision [33] 43] This scenario is consistent with the multidisciplinary effort urged to develop the new science of complex systems [44] It ....

....first derivative of a Gaussian) is recommended [47] In [47] no effective solution for combining detection properties of step edge and ridge operators was found. In computer vision literature, contour and texture detectors employ even and odd symmetric filters either computed as quadrature pairs [29], 30] 40] 49] or as complex Gabor functions [42] 48] 50] 51] It is known that Gabor functions provide band pass filters working in a joint space spatial frequency domain (such that a Gaussian frequency window band passes the spatial frequency content of an image area localized by a ....

[Article contains additional citation context not shown here]

W. T. Freeman, Steerable Filters and Local Analysis of Image Structures, Ph. D. Thesis, Massachusetts Institute of Technology, 1992.

....finder , we need an oriented wavelet transform. In addition, instead of a raw wavelet transform of an image, we measure the orientation strength along a particular direction 2 by the squared output of the even and odd phase quadrature pair of bandpass filters steered to the angle 2 [11]. This spectral power is called the oriented energy E(2; s; x; y) and E(2; s; x; y) h G 2 (s; x; y) i 2 h H 2 (s; x; y) i 2 (9) where G 2 (s; x; y) and H 2 (s; x; y) are the outputs of a quadrature pair of analyzing wavelet filters at the scale space coordinate (s; x; y) ....

....more difficulties. A very important property of a saliency detection scheme is whether it is invariant to similarity transformations (rotation, translation, scaling) The wavelet transform is a linear transform and is itself invariant to these transformations [8] In the wavelet transform we use [11,30], the smoothed signal at each scale is sampled but its wavelet transform is not. These wavelets do not produce a completely translation invariant representation. Since all other operations are linear (except that the wavelet coefficients are squared) this scheme is globally invariant for normal ....

[Article contains additional citation context not shown here]

W. T. Freeman, Steerable Filters and Local Analysis of Image Structure, Ph.D. Thesis, Media Arts and Sciences Section, School of Architecture and Planning, M.I.T., June 1992.

....by noise cancellation techniques were applied to the synchrosqueezed plane, using the most efficient techniques available [18] these also failed to produce any useful summaries. ffl Iterated pairing. This method uses dynamic programming to implement saliency extraction by iterated pairing [20, 19, 21, 6]. Structures are built up locally by iteratively pairing together substructures. For the synchrosqueezed plane images, this method turned out to be very unreliable. Apart from being slow to implement, it does not have the temporal smoothness bias built in, and tends to find random structures ....

W. T. Freeman. Steerable filters and local analysis of image structure. PhD thesis, M.I.T, Cambridge, MA, June 1992.

....addition, the computation performed on the second stage is novel because it is probably global and is guaranteed to find the optimal axis. Dynamic programming has been used before but only on bitmap images and without probably global results (due to the need for smoothing or curvature computation) [3] [13] 14] 8] These schemes use non global approximations of smoothness base on energy minimization or curvature propagation. The approach presented here is an extension of Curved Inertia Frames as presented in [9] The current extension makes the scheme probably global and introduces the notion ....

W.T. Freeman. Steerable filters and local analysis of image structure. PhD Thesis. M.I.T. 1992.

....is the use of one dimensional operators. These methods take a previously determined edge map, and process this in some way (e.g. differentiate it twice) to give a measure of the amount of curvature present (e.g. 30] 6] A slightly different approach has been taken by, for example, Freeman [12] or Anderson [1] in which steerable filters are used to obtain information about possible multiple orientations in one region, indicating the presence of complex neighbourhood features such as corners. Also, a scale based method was proposed by Rattarangsi and Chin [29] These methods are still ....

W. T. Freeman. Steerable Filters and Local Analysis of Image Structure. PhD thesis, Massachusetts Institute of Technology, USA, 1992.

.... = I2(x) Theta (1 Gamma M(x) I2 (x) 1 Gamma M(x) Theta = I(x) d) I1(x) Theta M(x) I1(x)M(x) Theta = I2(x) Theta (1 Gamma M(x) I2(x) 1 Gamma M(x) Theta = I(x) Figure 2: Kinetic occlusion: Layers model (a,c) where x = x; t) b,d) where x = x; y; t) junction analysis [11, 16, 31] to space time. By detecting motion boundaries, other visual modules may be aided: motion estimation can be improved by indicating where smoothing of motion estimates should not occur; and, as spatiotemporal junction analysis yields information similar to spatial junction analysis, our module can ....

....our kinetic occlusion detection architecture. First, considerable improvements can be made on the details of our detection methods alone. Other mechanisms which signal the presence of a given orientation and motion can be easily constructed. Combining techniques used in spatial junction analysis [11, 16, 31] and our spatiotemporal junction analysis in a unified framework is in order. Second, we are currently using non causal operations. An asymmetry in detecting appearance and disappearance is expected; considering the temporal dynamics of distributed winner take all mechanisms could handle this ....

W. Freeman. Steerable filters and the local analysis of image structure. PhD thesis, MIT Dept. Media Arts and Sciences, September 1992.

No context found.

W. T. Freeman. Steerable Filters and Local Analysis of Image Structure. PhD thesis, Massachusetts Institute of Technology, Cambridge, MA, 1992. URL: http: // citeseer. nj. nec. com/ freeman92steerable. html

No context found.

W. T. Freeman. Steerable Filters and Local Analysis of Image Structure. PhD thesis, MIT Media Lab. TR-190, June 1992.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC