Fourth International Workshop on Advanced Computational Intelligence Wuhan, Hubei, China; October 19-21, 2011
A New Circle Targets Extraction Method from High Resolution Remote Sensing Imagery Z. Chen , J.G . Liu , and G.Y. Wang
application. The Hough transform algorithm is proposed by Paul Hough in 1962, which can detect the arbitrary analytic curve in the image space[1]. Now, the Hough Transform (HT) has been a standard method for shape recognition in digital images. It was first applied to the recognition of straight lines[2] and later extended to circles[3][4][5], ellipses[4] and arbitrarily shaped objects[2][6]. The circle Hough Transform method in the [3] proposed by Davies is applied widely in the circle object detection because it is able to complete the detection task under the noise, distortion condition. The traditional circle Hough transform needs 3-dimension parameter space and when the parameter space exceeds two dimensions, the computing time and the storage requirement of the Hough transform (HT) will increase dramatically. The traditional HT treats all angles equally, which also results in heavy computation, huge parameter space, and less salient peaks. Various literatures were proposed to modify the conventional HT. They mainly focus on how to select angles to conduct the transform accurately and efficiently. HT is insensitive to missing parts of lines, to noise, and to other non-line structure co-existing in the image, and it may search for several curves in one pass of the process, however, HT suffers several flaws [7]: (1) Huge computation and memory storage. (2) Low speed. (3) Depending on the number of parameters and the split of the parameter space. (4) Difficulties in finding local maxima if peak is not properly defined. A lot of research has been carried out in effort to look for an effective and convenient Hough Transform circle detection technique. In order to avoid the waste of large amount of memory space caused by the parameters accumulation, Chen et al. [8] presented the randomized circle detection (RCD). But these methods have to depend on the neighborhood information of the sampling point, which leads to a weak robustness, or their computational complexity is high. In [3], Davies made use of the circle object geometrical feature that the line passes the center of a circle if the straight line is vertical to the tangent of the point on the circle boundary to detect this circle object. But the detection is susceptible to noise since the gradient direction's accuracy is often not high. Several methods utilize randomized selection of edge points and geometrical properties of circle instead of using the information of edge pixels and evidence histograms in the parameter space. However, the randomized selection method has its own problems such as probability estimation, accuracy and speed that are dependent on the number of edge pixels[9].
Abstract—According to the drawback of the traditional circle target extraction algorithm from high resolution remote sensing imagery used by Hough Transform, such as computation complexity, low efficiency and etc, a new circle target extraction method is proposed in this paper which can extract multiple circle targets with different radius at one time. First, the Average Absolute Difference is implemented to enhance the edge of the circle targets and suppress the noise of the background. Secondly, the locally self-adaptive segmentation algorithm is implemented to obtain the binary image. Thirdly, the thinning algorithm based on model computation is implanted to obtain the single pixel edge of the circle targets and in order to reduce the computation times in the following process. Furthermore, a pruning algorithm is necessary; finally, a modified Hough transform algorithm is proposed to obtain the circle targets. The experimental results demonstrate that the new circle targets algorithm can extract the multiple circle targets quickly and accurately, which has three advantages: low time consuming, high detection rate, robust to noise and fragmentary boundaries.
I. INTRODUCTION
H
igh-resolution remote sensing imagery like IKONOS and QuickBird has been applied in many fields in recent years. The main difference between high-resolution remote sensing imagery and low-resolution or medium-resolution remote sensing imagery such as TM or SPOT is that high spatial resolution remote sensing imagery provides more useful information than low resolution or medium resolution remote sensing imagery. But it brings about more noise which can affect the recognition of the targets also. Circles are a common geometric structure of interest in computer vision applications. It is an important work of image analysis to extract their position in the image accurately and effectively. The Hough Transform and its modified algorithms are the tool used usually in the Manuscript received Jun 21, 2011. This study was supported in part by the Nature Science Fund of China for Young Scholars under grant No. 40801164 and Defense Innovation Fund of Huazhong University of Science and Technology. Z. Chen is with the Intelligence Control, Institute for Pattern Recognition and Artificial Intelligence, Huazhong University of Science and Technology and Multi-Spectral Information Processing State Key Laboratory (corresponding author to provide phone: 086-027-87557746; fax: 086-027-875577465; e-mail:
[email protected]). JG Liu is with the Intelligence Control, Institute for Pattern Recognition and Artificial Intelligence, Huazhong University of Science and Technology and Multi-Spectral Information Processing State Key Laboratory. GY Wang is with the Intelligence Control, Institute for Pattern Recognition and Artificial Intelligence, Huazhong University of Science and Technology and Multi-Spectral Information Processing State Key Laboratory.
978-1-61284-375-9/11/$26.00 @2011 IEEE
529
In this paper, we propose an effective and convenient circle detection method based on modified Hough Transform. The method not only fully utilizes the circle target's geometrical feature but also reduces the computing times deeply, the accuracy position of the circle targets’ edge are obtained also. The new circle targets extraction method include four steps mainly. First, the a non-linear filter and Average Absolute Difference is implemented to suppress the noise of the background and enhance the edge of the circle targets. This step process helps preserving the connectivity of the circle targets’ edge also. Thus, the whole and closed boundaries of the circle targets will be obtained with the followed algorithm. Second, in order to obtain the binary image, a locally adaptive thresholding algorithm was utilized to segment the circle targets from the background. Thirdly, because the boundaries of the circle targets obtained by the second step are not the single pixel, a thinning algorithm is implemented. The thinning algorithm is implemented to remove the short branch in the skeleton image. Finally, the modified Hough Transform is proposed to obtain the circles’ position. Section 2 describes the methodology of this new method. Section 3 presents an example of the new method, and Section 4 concludes the paper. II.
determining the property of average absolute difference based on their neighboring pixels in natural scenes In many cases, the target is conspicuous in a local region and the brightness of the target is different from that of the background. Thus, the Average Absolute Difference Maximum is proposed according to the correlation between the pixel and the neighborhood or the smoothness of the space domain. This also means that the target is conspicuous in a local region. Taking this fact in account, we use a double-window filter to enhance the target. Set Θ and denote the pixels in the internal window and pixels between internal window and external window, respectively. If the size of external window is slightly bigger than that of the object and internal window can be changed, then the AADM can be defined as[12][13]:
' PD[ PD[ ' _ ' 1
,
[
[
1
,
\
\
˄˅
where N and N denote the numbers of pixels in set Θ and , I x and I y denote the levels of pixel x and y , respectively. The object and the scene window are showed in fig. 1.
METHODOLOGY
This section details the image pre-processing, Average Absolute Difference, adaptive thresholding segmentation algorithm, thinning algorithm, thinning algorithm and modified Hough Transform for the circle targets extraction method from the high-resolution remote sensing imagery data.
Fig.1The object and the scene window
A. Noise Removal Noise removal and image smoothing are important to many images processing application. One of the drawbacks of traditional filters (e.g. Gaussian and median-filtering) is that typically they are applied uniformly across the image—modifying pixels that are not corrupted by noise and leading to errors in the final results. To circumvent this problem, a nonlinear ‘‘peer group filtering’’ (PGF) [10][11] noise filter was proposed. The main steps of this filtering procedure are Peer group size selection, Peer group classification and Pixel replacement. The peer group size is the number of the peer group members implemented to smooth the center pixel of the local window. The purpose of averaging over the peer group members instead of the entire local window is to avoid edge and details blurring. The peer group classification is to decide that the center pixel of the local window belongs to the impulse noise and does not have the peer group. The pixel replacement is to replace the center pixel of the local window with Gaussian weighted average of its peer group members.
When includes object and the includes background only, the dissimilarity degree will be maximum. So according to the conclusion, the objects’edge will be obtained and enhanced. The result is shown in Fig.4. C. Local Thresholding
After the edge enhanced image is obtained, the quality of the image is improved. However, the image is still a gray image and the edge of the objects is still faint white. So, to obtain a better representation of the shape of the circle targets, it is necessary to separate the circle targets from the image background. Due to the fact that the gray-level intensity values of the circle targets vary at different locations in the image, global threholding techniques do not provide satisfactory results. Hence, a locally adaptive thresholding segmentation algorithm based on local histogram was utilized to segment the circle targets patterns from the background. The algorithm can choose different threshold values for every pixel in the image based on the analysis of its surrounding neighbors. If f ( x, y ) characterizes the spectral values of pixel centered within a square window having dimension n and T ( x, y ) denotes the local threshold value in the square
B. Average Absolute Difference Maximum As originally conceived, the problem addressed in this paper is to enhance the objects’ boundary and remove the noise from natural scenes. In essence, this involves
530
window. The segmentation result is obtained through the following formula:
f ( x, y ) 0 f ( x, y ) 255
f ( x, y ) T ( x, y ) f ( x, y ) T ( x, y )
cn ( p)
1 8
pi pi 1 2 i 1
(3)
8
(2)
sn ( p) pi
(4)
i 1
The result is showed in Fig. 5;
cn ( p) and sn ( p) are the number of the cross and skeleton points in the neighbor respectively. pi are the gray Where
D. Thinning and Pruning
For the circle target, the shape is most important factor. A good representation of the pattern’s shape is via extracting its skeleton. And also, through extracting the skeleton, the numbers of the pixels which need be computed and the computation will be reduced. The skeleton with the single pixel edge image is obtained by the thinning algorithm that reduces the width of lines to one pixel. The algorithm of skeletonization performs iteratively in 8 subsequent steps, where each step peels one layer of pixels in a certain direction. As one iteration is accomplished, the pattern is one pixel thinner in all directions. The algorithm stops when no difference between the input and the output is obtained[14][15].
values of the neighbors and valued as 0 or 1. If the pixel is the skeleton point, pi is value as 1. Otherwise, pi is value as 0. The judgment condition of the skeleton point’s type is defined as: (1) If cn ( p) s n ( p) 1 , skeleton point is end point. (2) If
cn ( p) sn ( p) 3 , skeleton point is end point.
(3) Other conditions, the point is common skeleton point. It can be seen that after the pruning process, the skeletons of the objects are successfully extracted and the shape of the objects is well preserved. The process helps reducing the amount of calculation in the followed step also. The thinning and pruning results are showed in Fig.6 and Fig.7 respectively. E. Modified Hough Transform
The Hough transform can be used to determine the parameters of a circle. A circle with radius R and center (a, b) can be described with the parametric equations[16]
x R cos( ) a y R sin( ) b
(5)
The traditional Hough Transform is to transform the feature points into the parameter space. When the angle θ sweeps through the full 360 degree range the points (x, y) trace the perimeter of a circle. If an image contains many points, some of which fall on perimeters of circles, then the job of the search program is to find the parameter triplets (a, b, R) to describe each circle. The fact that the parameter space is 3D makes a direct implementation of the Hough technique more expensive in computer memory and time. So, a modified Hough Transform is proposed in this paper. After the edge points ( xn , yn ) is mapped into the parameter space, the corresponding coordinate point sets in the parameter space is obtained. Fn {( xn , yn ) | ( xn x0 )2 ( yn y0 )2 R 2} (6)
Fig.2 Different Skeletonization Templates Corresponding To The Direction Of “Peeling”.
Though the skeletonization is obtained by the thinning algorithm, the unwanted short branches and little area are removed by the pruning. In this paper, the pruning is achieved by the region growing algorithm. For the little area removal, after the area is obtained through region growing, if the size of the objects is less than the size threshold preestablished, the corresponding region will be removed. For the short branch of the object, the pruning is achieved by the region growing algorithm also. First, the end point and the cross point of the short branch is obtained through calculating the relationship between the neighbors. Secondly, the length between the end point and the cross point is calculated. If the length is less than a threshold, all the points in this short branch are removed. Through the two steps all
When the centre of a circle ( x0 , y0 ) is translated to the point ( x0 m, y0 n ), the every coordinate points of the circle ( xn , yn ) will be changed into ( xn m, yn n ). Thus, the coordinate point sets that the every edge points are projected in the parameter space do not need be calculated, which only need calculate the projected coordinate point sets Fbase of a reference point, the
the short branches are removed. In order to extract the end point and cross point of the branch, for the every pixel in the skeleton image, the numbers of the cross and its neighbors which are skeleton points need be calculated as follow.
531
projected coordinate point sets of the rest edge points can be calculated through translating the sets Fbase . Fbase {( xbase , ybase ) | ( xbase x0 ) 2 ( ybase y0 ) 2 R 2 } (7) The new modified Hough Transform reduces the computation times and improves the detection accuracy. The main steps are outlined as follows: (1)Select a reference point ( xbase , ybase ) in the initial
image, the projected coordinate points sets are obtained by projecting the edge points of the circle with the center ( xbase , ybase ) and the radius R in the parameter space . (2) Traversal the whole image and finding the first edge point ( x1 , y1 ) ; Adding the offset x1 xbase , y1 ybase to
Fig.5 the segmentation result Fig. 6 the thinning result
the all element of the Fbase , thus the new coordinate points sets F1 . Then, accumulate corresponding points in the parameter space of F1 . (3)Repeat the step (2) and calculate the Fn (2 n N ) ,
Fig.7 the pruning result
Fig.8 the final extraction result
There are four circle targets with different radiuses in the initial image. Through the processing by the algorithm proposed in this paper, every positions of the circle are detected accurately in the final extraction result. In order to illustrate the advantage of the computation speed with the new algorithm proposed in this paper, a table is used to compare the run time of the circle extraction between the traditional Hough Transform and the new modified Hough Transformation proposed in this paper.
N is all the number of the edge, while all the number of edge points are processed. (4)the accumulation numbers in the parameter space exceeds the threshold number set beforehand then the coordinate point in the parameter space is the coordinate point in the image space. So, the center of the circle with radius R is detected; The above steps is to improve the calculation speed mainly with a fixed radius R. In order to extract different TableI radius, a loop between the minimum possible radius and The Run Time Comparison between the New Algorithm(NHT) and the maximum possible radius is done. Traditional Hough Transformation(THT) Through all the process, the multiple accuracy center Method Times point coordinates of circle with different are extracted at NHT 0.8s one time quickly. The test result is show in Fig.8. THT 10s Through the comparison, the run time used by the algorithm proposed in this paper is more less than that of the traditional method.
III. RESULTS The image selected in this paper is 1 meter resolution. The image Size is 628h563. The test is done by the PC with 2.1GHz Inter P4 CPU and 1G EMS memory. The Fig.3 is the initial image. The Fig.4 shows the results through the Average Absolute Difference Maximum process. The Fig. 5 describes the segmentation result and the thinning and pruning results is showed in Fig.6. The final extraction result with modified Hough Transform is displayed in Fig.7.
Fig. 3 the initial image
IV. CONCLUSION Through the above results, the method proposed in this paper can not only detect multiple circles accurately but also improve the run time deeply. The algorithm proposed in this paper reduces the run time through the two processes. The number of the edge points is reduced through edge enhancement and the pruning algorithm at one side, on the other side, the modified Hough Transform is achieved through translation in the parameter space. So, the usage with the two algorithms can reduce the run time effectively. Because the edge enhancement algorithm can extrude the circle edge and enhance the edge position and the modified Hough Transform inherits the advantage of accuracy location, it can detect the circle position precisely.
Fig.4 the enhanced image result
532
ACKNOWLEDGMENTS This study was supported in part by the Nature Science Fund of China for Young Scholars under grant No. 40801164.
REFERENCES [1] [2]
[3] [4]
[5]
[6]
[7]
[8] [9]
[10]
[11]
[12]
[13]
[14]
[15] [16]
P. V. C. HOUGH, “Method and means for recognizing complex patterns”, U. S. Patent 3, 069 654, December 18 , 1962. R.O. DUDA and P. E. HART, “Use of theHough transform to detect lines and curves in pictures”, Communications of the ACM, 1972, vol. 15, no.1, pp.11-15. E. R. DAVIES, “A modificd Hough scheme for general circle location”, Pattern Recognition Letters, 1988, 6, pp.37-43. A. GONEID, S. EL-GINDI , and A. SEWISY . “A method for the Hough transform detection of circles and ellipses using a 1dimensional array”,Proc of the IEEE International Conference on Systems, Man and Cybernetics. Orlando, 1997, pp.3154-3157. IIRD P. KICRKEGA “A method for detection of circular arcs based on the Hough transform”, Machine Vision and Adulications,1992,5,pp. 249-263. D. J. KERBYSON, T. J. ATHERTON . “Circle detection using Hough transform filters”, Washington DC: IEEE Conference Publication, 1995:pp.370-374. L. Xu, E. Oja, and P. Kultanen, "A new curve detection method: randomized Hough transform (RHT)," Pattern Recognition Letters, vol. 11, no. 5, pp. 331-338, 1990. T. Chen and K. Chung, An efficient randomized algorithm for detecting circles, 2001 Ali Ajdari Rad, Karim Faez, Navid Qaragozlou. “Fast Circle Detection Using Gradient Pair Vectors, ” Proc. VIIth Digital Image Computing: Techniques and Applications, 10-12 Dec. 2003, Sydney, pp.879-887. Y. Deng , C. Kenney , M. S. Moore , “Peer Group Filtering and Perceptual Color Image Quantization,” 1999, Proceedings. IEEE International Symposium on Circuits and Systems, 4: 21-24. C. Kenney , Y. Deng, B. S. Manjunath, and G. Hewer. Peer Group Image Enhancement. 2001. IEEE Transactions on Image Processing. 10(2):326-334. Zhenxue Chen, Guoyou Wang, Jianguo Liu, Chengyun Liu, “Infrared small target detection algorithm based on self-adaptive background forecast”, International Journal of Infrared Millimeter Waves 27, pp. 1619–1624, 2006. G.Wang, T. Zhang, L. Wei. “Efficient method for multiscale small target detection from a natural scene”, Opt.Eng , vol. 35, no. 3, pp. 761–768, 1996. Suleyman Malki, Yu Fuqiang, Lambert Spaanenburg. “Vein Feature Extraction Using DT-CNNs”, The 10th International Workgroup on Cellular Neural and Their Application, Istanbul, Turkey, pp. 28-30 August 2006. C.Y. Suen and T.Y. Zhang, “A Fast Parallel Algorithm for Thinning Digital Patterns”, Communications of the ACM, vol. 27,no. 3 , 1984 Digital Image Processing, C. G. Rafael and E. W. Richard, Addison-Wesley Publishing Company,1992.
533
