Line and Edge detection

Line and Edge Detection

Introduction

In the past 20 years[1], many approaches have been developed to deal with the detection of linear features on optic or radar image. Most of them combine two criteria: a local criterion evaluating the radiometry on some small neighborhood surrounding a target pixel to discriminate lines from background and a global criterion introducing some large-scale knowledge about the structures to be detected.

Line detection can be used for several applications, such as registration with other sensor images, cartographic applications, and geomorphologic studies.

Two criteria for optimization are considered: 1) maximizing the total probability of detecting an edge within a window and 2) maximizing the accuracy with which the edge position can be determined[30].

The general principles of segmentation, of which edge detection can be part, were outlined by Oliver [31].

Ideal edge detection would provide both the greatest probability of detecting the presence of an edge within a test window, and also the most accurate edge-position determination[30]. The polarimetric whitening filter (PWF) optimally processes the polarimetric scattering matrix into pixel intensity in such a way as to reduce the speckle content of a SAR image [39].

Lee's filtering method with eight edge-aligned windows preserves the edge features and smoothes the homogeneous areas

[40]

.

Method 0: Line and Edge Detector

Main idea: The operator based edge detection methods have their shortcomings in that they only represent the local edge features in the images. Low threshold for the binary decision will result in a correct and continuous boundary, but at the expense of a high false alarm rate; while high threshold may lead to incomplete edges, despite a low false alarm rate[38].

Edge detector: Roberts, Sobel, Prewitt, Krisch, Gaussian smooth filter, Laplace Sharp filter, Gaussian Laplace filter, ratio average detector, morphologic edge detection, Canny, Zero-cross, Log-normal,

Paper: Existing methods were mostly applied on single-channel multi-look intensity images and differ by the user comparison criterion [25]. The most well-known is the ratio detector [13] which considers the ratio of average intensities between the two rectanglesand widely used in coherent imagery.

Paper: The ROA operator was proposed by Ridha Touzi [13], the GLR operator was proposed by Oliver [30],the coastline detection was proposed by J. Lee [41], the Duda operator was proposed by G. Geling and proved to be quite effective [42].

1

Paper: CFAR edge detector [13] is defined.

Method 1: Dynamic Programming

Main idea: Dynamic programming is used to minimize some global cost function, as in the

[2][3].

original algorithm of Fishlerand its improvements

Paper: It has also been applied on SAR images

[4] and [5]

.

Method 2: Hough Transform

Main idea: Most previous studies of road detection from remotely sensed images use space filters or edge filters which evaluate local characteristics assuming a road feature is distinctive in a window [27, 28]. The Hough Transform approach, on the other hand, can handle more global large line-like features in theory. However, only a few studies of line detection from SAR images have been done on very small test images .

Paper: The Hough transform is a coordinate transform of an image in which a straight line in the x-y plane is commonly projected to a single point in the transformed R-theta plane . Thus the problem of detecting lines in the image reduces to that of detecting points in a transformed Hough space. Peak heights of these points correspond to the line lengths. In an actual image, there are many other pixels which do not belong to lines and all of them also appear as noise in the Hough image. Furthermore, lines in a SAR image sometimes disappear, wiggle, or attach to neighboring objects and therefore the corresponding peaks in the Hough space fade irregularly.

Paper: Hough-transform-based approached has also been tested for the detection of parametric curves, such as straight lines or circles

Paper: Location of road intersection was estimated automatically from ERS-1 single look SAR imagery using a Hough Transform based approach

[26]

[5]-[7].

[29]

[29]

.

Method 3: Tracking Methods

Main idea: Tracking methods are another possibility. They find the minimum cost path in a grapy by using some heuristics, for instance, an entropy criterion[8].

Paper: A new method is proposed for joint detection of roads in multi-band SAR images[15]. First, the multi-segmented poly-line model is introduced to provide a more accurate description of road curve. Then, the roads in SAR images are extracted in a Bayesian tracking framework, and the particle filtering algorithm is employed to implement the tracking. Finally, a joint detection method is proposed to determine the optimal weights of particles based on the maximum likelihood criterion.

Paper: Mckeown and Denlinger[45] proposed a road-tracking algorithm for aerial images, which relied on road-texture correlation and road-edge following. This algorithm is semi-automatic.

2

Method 4: Energy Minimizing Curves

Main idea: Energy minimizing curves, such as snakes, have also been applied[9][22].

Paper: Gruen and Li [46] also proposed a semi-automatic road extraction algorithm for aerial images. They used the least squares B-spline snakes (LSB-Snakes) algorithm in multi-image mode, which provided a robust and mathematically sound 3-D approach.

Method 5: Bayesian Framework

Main idea: The Bayesian framework, which is well adapted for taking some contextual knowledge into account, bas been widely used.

Paper: Regazzoni defines a cooperative process between three levels of a Bayesian network, allowing the introduction of local contextual knowledge as well as more global information concerning straight lines[10].

Paper: Hellwich[11]uses prior information concerning line continuity expressed as neighborhood relations between pixels.

Paper: Florence Tupinproposed a new formulation using segment-sites. Since our aim is to detect the major roads present in an image, contextual knowledge on the scale of pixels is insufficient and results in numerous, small, disconnected road segments. However, on the scale of segments a few pixels long, a priori knowledge allow for the detection of the main axes in the road network. In the first step, road-segment candidates are detected. In the second step, a graph of segments is built and a novel Markov random field is defined to perform road detection, thus providing a new approach.

Paper: A novel road detection method based on Bayesian tracking framework is proposed [23].

[1]

Method 6: Statistical Properties

Main idea: In the case of radar imagery, local edge or line detectors are often based on statistical properties [12] or on the intensity ratio of neighboring regions [13], [14].

[32]

Paper: Frost applied two different hypothesis tests: one for homogeneity within a window, the other for a specific edge between two regions. The latter introduced more prior knowledge and hence gave better noise rejection at the expense of more complicated processing.

Paper: A second approach [33, 34] tested whether the two halves of a window had a different mean value in terms of the difference of the means normalized by their standard deviation.

Paper: A measure base on the ratio of the mean intensities in the two regions has also been proposed [13, 35].

Paper: This can be shown to provide maximum-likelihood total-edge-detection probability when the regions are of equal size [13, 36].

Paper: If the regions are not of identical size, no equivalent estimator has been postulated though an approximate solution, based on the Student t test, has been adopted [37] in a test to

3

determine whether to merge dissimilar-sized regions.

Paper: There have recently been a number of constant rates of false alarm edge detectors for SAR imagery developed as summarized in [43] and [13].

[44]

Paper: The combination of anti-parallel lineshas been used to develop a linear feature detector with a low probability of false alarm.

Paper: Geometric-probabilistic models were first built for road-image generation, and roads were found from it using MAP estimation. Barzohar and Cooper

[47]

presented an automated

approach to locate the main roads in aerial images. Tupin[1] proposed a nearly automatic detection algorithm for linear features such as the main axes of road networks. They presented two local line detectors as well as a method for fusing information from these detectors to obtain segments. The real roads were identified among the segment candidates by defining a Markov random field for the set of segments. Jeon[48] proposed an automatic road detection algorithm for satellite images. They presented a map-based method based on a coarse-to-fine, two-step matching process. The roads were finally detected by applying snakes to the potential field, which was constructed by considering the characteristics and the structures of roads.

Method 7: Pattern Recognition

Paper: Genetic algorithm detection in SAR images.

[16]

and morphological characteristics

[17]

were used for road

Method 8: Particle Filter

Main idea: Particle Filter (PF), which is a sequential Monte Carlo method based on the concept of importance sampling and the utilization of Bayesian theory, has captured the attentions of many researchers in various areas including signal processing , statistics, and econometric. This interest stems from the potential of particle filter for coping with nonlinear and\\or non-Gaussian problem, which is a tough task for traditional filtering methods.

Paper: recently, inspired from the snake method [18], Chen [19] has applied the particle filter to track consecutive road segments in the Bayesian tracking framework. The particle filtering algorithm [20-21] is a sequential Monte Carlo method based on the concept of importance sampling and Bayesian theory.

Reference Papers

[1] Florence Tupin. \Extraction.\

[2] M. A. Fischler, \multisource knowledge integration technique\

[3] N. Merlet,\

4

[4] R. Samadani and J. F. Vesecky. \[5] J. W Wood. \

[6] S. Quegan. \

[7] J. Skingley and A. J. Rye.\1987.

[8] D. Geman and B. Jedynak. \active testing model for tracking roads in satellite images\1996.Load

[9]P. Fua and Y. G. Leclerc.\

[10] C. S. Regazzoni. \adaptive probabilistic model for straight edge-extraction within a multilevel MRF framework\

[11] O. Hellwich. \[12] C. J. Oliver. \

[13]Rodha Touzi. \ Statistical and Geometrical Edge Detector for SAR images\

[14] M. Adair and B. Guindon.\SAR images\

[15]Qiming Deng. \based approach to road detection in multiband SAR images\2009.Load

[16] B. Jeon, J. Jang, and K. Hong.\in space-borne SAR images using a genetic algorithm\

[17] C. Zhu. %using image morphological characteristics\

[18] L. Bentabet. \[19] Y. L. Chen. \[20] A. Doucet. \

[21] J. S. Liu. \[22]M. Kass. \

[23]Yilun Chen, \

[24] D. Borghys. \[25] R. Fjortoft. \

[26]JojiIisaka. \[27] F. Wang. \ knowledge-based system for highway network extraction\[28] L. Alparone. \

[29] R. O. Duda and P. E. Hart.\1972.

[30] C. J. Oliver. \[31] Oliver. \

[32]Frost. \[33] White, R. G. \[34] White, R. G. \[35]Bobik, A. C. \

[36] Caves,R. \matching of features in synthetic aperture radar data to digital map data\

[37] Cook, R. \[38] G. Y. Zhou. \ new edge detection method of polarimetric SAR images using the

5

curvelettransform and the Duda operator\

[39] L. M. Novak. \

[40]J. Lee. \[41] J. Lee. \

[42] G. Geling. \[43] A. Lopes. \

[44] R. Nevatia and K. Ramesh Babu.\

[45] D. M. McKeown and L. Denlinger. \methods for road tracking in aerial imagery\

[46] A. Gruen and H. Li. \multiple images by 1sb-snakes\1997.

[47] M. Barzohar and D. B. Cooper.\aerial images by using geometric-stochasic models and estimation\

[48] B. Jeon. \

Reference Authors

Florence Tupin M. A. Fischler Rodha Touzi C. J. Oliver Frost J. Lee

Reference Algorithms

Scheme 1: multi-segmented poly-line road model + Bayesian tracking framework + Particle filter + Joint detection of roads based on the particle filter[15].

Scheme 2: tracking framework + Bayesian Filter + Particle Filter [23].

Scheme 3:multi-variate + ratio detector [24].

Scheme4: small feature preserving filtering method

[28]

+ fuzzy membership function +

Hough Transform + morphological filtering + overlaying images[26].

[30]

Scheme 5: hypothesis testing + Student t test + Touzi-ratio + ………………………… Scheme 6:curve-let transform + Duda operator + PWF + Lee Filter + threshold selection [38].

Scheme 7:Duda operator .

Scheme 8:Laplacian-of-a-Gaussian (LoG) + ratio-of-average (RoA) + zero crossings + speckle models + power spectral density .

Scheme 9: gradient edge detector + ratio edge detector + edge orientation + CFAR edge detector [13].

Scheme 10: extraction of curvilinear structures + perceptual grouping factors + genetic algorithm (GA) [16].

6

[35]

[42]

curvelettransform and the Duda operator\

[39] L. M. Novak. \

[40]J. Lee. \[41] J. Lee. \

[42] G. Geling. \[43] A. Lopes. \

[44] R. Nevatia and K. Ramesh Babu.\

[45] D. M. McKeown and L. Denlinger. \methods for road tracking in aerial imagery\

[46] A. Gruen and H. Li. \multiple images by 1sb-snakes\1997.

[47] M. Barzohar and D. B. Cooper.\aerial images by using geometric-stochasic models and estimation\

[48] B. Jeon. \

Reference Authors

Florence Tupin M. A. Fischler Rodha Touzi C. J. Oliver Frost J. Lee

Reference Algorithms

Scheme 1: multi-segmented poly-line road model + Bayesian tracking framework + Particle filter + Joint detection of roads based on the particle filter[15].

Scheme 2: tracking framework + Bayesian Filter + Particle Filter [23].

Scheme 3:multi-variate + ratio detector [24].

Scheme4: small feature preserving filtering method

[28]

+ fuzzy membership function +

Hough Transform + morphological filtering + overlaying images[26].

[30]

Scheme 5: hypothesis testing + Student t test + Touzi-ratio + ………………………… Scheme 6:curve-let transform + Duda operator + PWF + Lee Filter + threshold selection [38].

Scheme 7:Duda operator .

Scheme 8:Laplacian-of-a-Gaussian (LoG) + ratio-of-average (RoA) + zero crossings + speckle models + power spectral density .

Scheme 9: gradient edge detector + ratio edge detector + edge orientation + CFAR edge detector [13].

Scheme 10: extraction of curvilinear structures + perceptual grouping factors + genetic algorithm (GA) [16].

6

[35]

[42]

本文来源：https://www.bwwdw.com/article/b7yw.html