Edge Detection of Riverway in Remote Sensing Images Based on Curvelet Transform and GVF Snake

Size: px

Start display at page:

Download "Edge Detection of Riverway in Remote Sensing Images Based on Curvelet Transform and GVF Snake"

Job Bailey
5 years ago
Views:

1 Proceedings of the 8th International Symposium on Spatial Accuracy Assessment in Natural Resources and Environmental Sciences Shanghai, P. R. China, June 25-27, 2008, pp Edge Detection of Riverway in Remote Sensing Images Based on Curvelet Transform and GVF Snake Moyan Xiao 1 +, Yonghong Jia 1, Zhibiao He 2 and Yan Chen 1 1 School of Remote Sensing and Information Engineering, 2 Satellite and Navigation Location Technology Research Center, Wuhan University, 129 Luoyu Road, Wuhan , P.R. China Abstract. This paper introduces curvelet transform and gradient vector flow (GVF) snake to improvement accuracy in edge detection of waterway from remote sensing images. Multi-scale geometric analysis (MGA) is boog hot research topic in recent years, which aims to obtain flexible, fast and effective signal processing algorithms through efficient approximation and characterization for the inherent geometric structure of high-dimensional data. Curvelet transform is a special member of this emerging family of MGA which overcomes inherent limitation of traditional multi-scale representation such as wavelet which ignores the geometric properties of objects with edges and does not exploit the regularity of the edge curves in higher dimension. The basic edge detection process is mainly composed of three parts. Firstly, obtain the initial snake based on region growing and morphology methods from curvelet-based denoised image. Secondly, get the edge map derived from curvelet-based enhancement image. Finally, obtain the converging snake by evolving the GVF snake. The edge detection results of Yangtze River derived from the proposed method, wavelet based GVF snake and canny method are compared together. Experiments demonstrate that the new algorithm is superior to other methods, which is more effective and accurate. Keywords: edge detection, remote sensing image, multi-scale geometric analysis (MGA), curvelet transform (CT), GVF Snake 1. Introduction The edge detection for remote sensing image is an important way to obtain the remote sensing information and the foundation for understanding the remote sensing images. Riverway course detection is an important part of hydrological monitoring. It can offer decision support for hydrological department to manage water resources, the flood disaster forecasting and so on. The accurate and efficient shoreline of waterway has also application in waterway GIS system [1, 2]. It is known that wavelet is one of the common scaling methods in Geo-science which is successful in the application in hydrology [3 ]and data fusion model [4] in remote sensing etc.. Though wavelet has good performance at representing point singularities, this is not the case in higher dimensions because wavelet ignores the geometric properties of objects with edges and does not exploit the regularity of the edge curves. For two-dimensional images, wavelet isolates well the discontinuities across the edges, but along the edges, it prevents the edges from being smooth. Multi-scale Geometric Analysis is boog hot research topic in recent years. The application based on MGA tools such as ridgelet, contourlet and curvelet [5] is successful in image fusion, detection, speckle reduction, identification, enhancement etc.. Curvelet transform [5-8] is a special member of this emerging family of MGA developed by Candès etc., whose approximate scales and orientations are supported by a generic 'wedge'. It owns very high directional sensitivity and anisotropy. It is more efficient for curvelet transform to represent edges and singularities along curves than that of the traditional wavelet transform. + Corresponding author. Tel.: address: xiao_mo_yan@163.com. ISBN: , ISBN13:

2 Fig.1. The flowchart for the edge detection of waterway algorithm Riverside detection is in the class of boundary detection problems. Similar problems have been encountered in papers such as coastline detection [9-11], riverway boundaries [12-13] as well as object recognition in various remote sensing images. They are most combined with wavelet-based edge detection and snake. For the wavelet-based snake is easy to be disturbed by the false strong edges and drift the snake into unwanted stage, an alternative method to detect riverside with reasonable accuracy is described to eliate the disturbance of false strong edges based on curvelet enhancement image in the paper. The proposed algorithm is mainly accomplished by three steps. Firstly, obtain the rough boundary based on region growing and morphology methods from curvelet-based denoised image. Secondly, get the edge map derived from curvelet-based enhancement image. Finally, obtain the converging snake by evolving the GVF snake. The band 5 TM image which covers Wuhan in China is used to validate the algorithm validity and Fig.1 presents the overall flow of the algorithm. 2. Curvelet Transform and GVF snake In this section, we present a brief review of the key theory to the algorithm involved, i.e. curvelet transform and GVF snake Outline of Curvelet Transform Curvelet transform is a new geometric multi-scale transform developed by Candès et al [5 8]. Curvelet transform decomposes the image into a series of high-pass and low-pass bands which is the same as wavelet transform. The wavelet transform extracts directional details that capture horizontal, vertical and diagonal activity. However, curvelet transform captures the structural activity along radial 'wedges' in the frequency domain. The approximate scales and orientations are supported by a generic 'wedge'. So curvelet transform owns very high directional sensitivity and anisotropy. The second generation curvelet transform [8] is introduced in the paper, which digital implementation is based on fast discrete curvelet transform (FDCT) i.e. the unequally-spaced fast Fourier transforms (USFFT) and the wrapping of specially selected Fourier samples. The wrapping algorithm is used in the paper. To evaluate it, firstly, apply the image with a 2-dimensional fast Fourier transform. Secondly, multiply it with scale and angle windows and wrap this product around the origin. Finally apply it with a 2-dimensional inverse fast Fourier transform. 345

2.2. GVF snake Snakes or Active contour models [14] are used extensively in computer vision and image processing applications, particularly to locate object boundaries.

3 2.2. GVF snake Snakes or Active contour models [14] are used extensively in computer vision and image processing applications, particularly to locate object boundaries. In GVF snake [15], the traditional external force is replaced by a new force field, called gradient vector flow, produced by a spatial diffusion of an edge map, which guides the propagation to boundaries from both sides. GVF snake, breaking through limit of initialization and poor convergence to boundary concavities of traditional snake, has a large capture range and is able to move snakes into boundary concavities. 3. Edge Detection of Riverway in Remote Sensing Images In this section, we propose a common edge detection model between land and water in remote sensing images. Aig to different source of the remote sensing images, more water bodies extraction methods may be considered to help the edge detection [16-18] Extracting the initial snake To obtain an initial snake, we first denoise the image by estimating the noise variance of the input image with the robust median operator [19].Secondly, compute the noise variance of each directional subband [20].Thirdly, and segment the land and water with seed or block [13] region growing method. Finally, extract the coarse contour by the morphology methods. For the sake of disturbance of noise, image should be denoised before region growing. For SAR image, the logarithm of the image is taken to transform the speckle to an additive noise. With the assumption that the same object has similar spectral characteristics, the pixel value of homogeneous water is close and different distinctly from the land. So we use region growing method to segment the water and land areas. As in the remote sensing image, even the same river may be splited by bridges, so after segmentation, the morphology methods of dilate, erode and edge extraction are used to obtain the initial snake. (a) (b) (c) Fig.2 (a) Original image. (b) Image of region growing segmentation. (c) Initial contour Obtaining the final contour Edge map is a key component of GVF snake, as GVF snake s external force is gradient vector flow field and GVF derives from edge map. We propose curvelet base enhancement method that not only enhances weak edge but also depresses noise efficiently. Because curvelet transform provides not only multiresolution analysis, but also geometric and directional representation. In the frequency domain, both weak edges and noises lead to low-value coefficients, however, weak edges different from noise have geometric structures; we can use this geometric representation to distinguish them. Moreover, as the proposed enhancement method enhances the weak edge at the expense of strong edge, weak riverway contour enhancement can help prevent snaxels (pixels of snake) over-drift. At a narrow river, snaxels are easy to be correctly located. It overcomes the shortcogs of other methods [10-12] whose snaxels are attracted by the strong waterway boundary. We enhance edges in an image through modifying the curvelet coefficients. In order to prevent from enlarging the noise and over enhancing the original clear edge, two thresholds are set. Let x denote the curvelet coefficient, and y(x) is edge enhance function which is similar to [21].The enhancement algorithm 346

4 in detail is as follows: y( x) = 1, x > T Tmax y( x) = x x T y( x) = T y( x) = 1or0, x max p,2t T T T max < x T p max 2T + T x, T < x 2T Here, p deteres the degree of nonlinearity. Usually, T is equal to k σ, and σ denotes the noise standard deviation. k is normalization parameter and larger than 3, which guaranties the noise will not be amplified. The parameter T max is the value under which coefficients are amplified. This value depends on the pixel values inside the directional subbands of different curvelet scale. Tmax may be derived from the maximum curvelet coefficient C max of each directional subband ( T max = lcmax, l < 1 ) or derived from the noise standard deviation ( T max = k m σ ) using an additional parameter k m. (a) (b) (c) Fig.3. (a) Enhanced image. (b)wavelet-based edge map. (c)new proposed edge map. The curvelet enhancement method for grayscale images consists of the following steps: (1) Estimate the noise standard deviation in the input image. (2) Perform a multi-scale decomposition of the input image using the FDCT. (3) Calculate the noise standard deviation for each directional subband in different scale of the curvelet transform (see [20] for more details on this step). (4) For each directional subband, calculate the maximum of the subband, which is used to detere T max. Then multiply each curvelet coefficient by y (x). (5) Reconstruct the enhanced image from the modified curvelet coefficients. After getting the enhanced image, the edge map can be obtained by using a 2-D Gaussian function to smooth it and process it with the gradient operator. Then, use the edge map to compute the GVF field. At last, evolve GVF snake to get the final snake using the foregoing initial snake and GVF field. 4. Experiment and Discussion Water bodies in middle infrared band such as Landsat TM band 5 have low spectrum reflectivity. But land and building etc. have high reflectivity. So we choose TM band 5 image (512*512pixels) covering Wuhan, China to extract contour of Yangtze River. To evaluate the effect of the proposed method, the results derived from the proposed method, wavelet-based GVF snake and canny method are compared to. In addition, the high resolution images corresponding to some representative zones are used to illustrate the proposed method validity. The original image, result of region growing segmentation and initial snake based on proposed method are displayed in Fig.2. The initial riverside has good continuity which avoids influence of the bridges. 347

(a) (b) 348 (c) (d) Fig.4. (a) Detected riverway contour by wavelet based GVF snake overlain on original image. (b) Enlargement of square area A in (a). (c) Enlargement of square area B in (a).

In order to get good river contour, the better wavelet based GVF snake edge detection method is first to obtain initial snake by [13] and the morphology methods.

In this experiment, as the original image has less noise, to obtain more edge information, the edge map is gained from sum pixel values of three high frequency local modulus maxima points.

5 (a) (b) 348 (c) (d) Fig.4. (a) Detected riverway contour by wavelet based GVF snake overlain on original image. (b) Enlargement of square area A in (a). (c) Enlargement of square area B in (a). (d)enlargement of square area C in (a). In order to get good river contour, the better wavelet based GVF snake edge detection method is first to obtain initial snake by [13] and the morphology methods. Then extract edge map according to [22] [11]. In this experiment, as the original image has less noise, to obtain more edge information, the edge map is gained from sum pixel values of three high frequency local modulus maxima points. Finally extract contour based on GVF snake. Enhanced image, edge map based on proposed method and wavelet-based edge map are shown in Fig.3. Fig.4 is the result based on wavelet based GVF snake after ten iterative operations. The finally detected contour in red is shown in Fig.4 (a). Fig4 (b), Fig.4(c) and Fig.4 (d) are enlargement images of square area A, B and C in Fig.4 (a) respectively. Fig.5 is the result based on the new proposed method after ten iterative operations. The final snake in red overlain on the original image is shown in Fig.5 (a). Fig.5 (b), Fig.5(c) and Fig.5 (d) are enlargement images of square area A, B and C in Fig.5 (a) respectively. To verify result further, we capture the high resolution image to depict the two methods difference. High resolution image according to square area A, B, C in Fig.4 (a) and Fig.5 (a) respectively and edge map based on canny method are shown Fig.6. Though, experiment results in Fig.6 (d) indicate that canny method has good edge detection effect, there are broken edges in riverway contour and some edges we don t want. Wavelet based GVF snake method and new proposed algorithm on the straight and wide apart shoreline can both overcome the bother of ships and get effective contour which can be seen from B and part of A in Fig.4 (a) and Fig.5(a) by contrast with high resolution images in Fig.6 (b) and (a).however, in the narrow river, the snake evolved by wavelet based GVF snake method is easy to be attracted by the other riverside which has stronger edge, which can be seen from A in Fig.4(a) and its enlargement Fig.4(b).In this aspect, the new proposed method effectively avoids the influence of strong edge by enhancing the wake boundary, which can be seen from A in Fig.5(a) and its enlargement Fig.5(b). Moreover, the new proposed method has high sensitivity to river with zigzag and is more excellent than wavelet based GVF snake which can be drawn from comparison among picture Fig.4 (d), Fig.5 (d) and Fig.6 (c).

image. (b) Enlargement of square area A in (a). (c) Enlargement of square area B in (a).

6 (a) High resolution image according to square area A in Fig.4 (a) and Fig.5 (a).

6 (a) (b) (c) (d) Fig.5. (a) Detected riverway contour based on the new proposed method overlain on original image. (b) Enlargement of square area A in (a). (c) Enlargement of square area B in (a). (d) Enlargement of square area C in (a). Fig.6 (a) High resolution image according to square area A in Fig.4 (a) and Fig.5 (a). (b) High resolution image according to square area B in Fig.4 (a) and Fig.5 (a). (c) High resolution image according to square area C in Fig.4 (a) and Fig.5 (a). (d)edge map based on canny method. 349

7 Furthermore, when we make an experiment on images contaated by an additive zero-mean Gaussian white noise, the results of the new proposed methodology is better than others. 5. Conclusion Cuvelet transform is well-adapted to represent edges. Effective image with noise reduction and enhanced edges can be obtained by reconstructing the modified cuvelet coefficients according to the proposed noise reduction method and enhancement function. The extraction of riverway contour is more effective and accurate by introducing curvelet transform and GVF snake into edge detection. In addition, the new proposed algorithm is superior to process noisy image and fast which is useful for real-time implementations. In a word, we are delighted to see the new emerging multi-scale transform curvelet transform is great promising in accurate improvement for remote sensing imagery interpretation. 6. Acknowledgements The authors thank all people for their useful discussions on image edge detection. 7. References [1] J Xi, ZL Xu, Design and Realization of Waterway Information Management System Based on GIS, Computer Technology and Development, 2006, 16 (1): [2] Q Wang, ZY Guo, JP WU,XY Gu, Design of Shanghai watercourse information system based on GIS, Shanghai Geology, 2006, 3: [3] CY TANG, R WANG, R MIAO, Application of Discrete Wavelet Transform in Hydrological Series Decomposition, China Rural Water and Hydropower, 2007, 2: [4] G Xiao, ZL Jing, L Herry, S Wang, A United Optimal Fusion Method of Pixel and Feature for Remote Sensing Images Based on the Statistical Properties of Wavelet Decomposition, Journal of Remote Sensing, 2005,9(4) : [5] E. J. Cand`es and D. L. Donoho, Curvelets a surprisingly effective nonadaptive representation for objects with edges. In C. Rabut A. Cohen and L. L. Schumaker, editors, Curves and Surfaces, Nashville, TN: Vanderbilt University Press, 2000: [6] E. J. Cand`es and F. Guo, New multiscale transforms, imum total variation synthesis: application to edgepreserving image reconstruction. Sig. Process., special issue on Image and Video Coding Beyond Standards, 2002, 82(11): [7] E. J. Cand`es and D. L. Donoho. New tight frames of curvelets and optimal representations of objects with piecewise-c2 singularities. Comm. on Pure and Appl. Math, : [8] E. J. Candès, L. Demanet, D. L. Donoho, and L. Ying, Fast discrete curvelet transforms, Multiscale Model. Simul, 2006, 5(3): [9] Mason D C and Davenport I J, Accurate and efficient deteration of the shoreline in ERS-1 SAR images. IEEE Trans. on Geoscience and Remote Sensing, 1996, 34 (5): [10] Niedermeier A, Romanee Ben E, and Lehner S, Detection of coastlines in SAR images using wavelet methods, IEEE Trans.on Geoscience and Remote Sensing, 2000, 38 (5): [11] LR Li, SX Gao; SF Cao, Detection of Shoreline in SAR Image Based on Wavelet and GVF Snake Model, Hebei Journal of Industrial Science & Technaology, 2004, 24(4): [12] JJ Zhu, HD Guo, XT Fan, Automatic and Fast Detection of Edges between Land and Water in High-resolution SAR Images, Remote Sensing Information, 2005, 5: [13] WB Wang, L SUN, XM Yi, PS Fei, A Wavelet Snake Method to Detect Boundaries Between Land and Water in SAR Images, Chinese Journal of Engineering Mathematics, 2007, 24(6): [14] M. Kass, A.Witkin, and D. Terzopulos, Snakes: Active contour models, Int. J. Comput. Vis., 1987, 1(4): [15] CY. Xu and J. L. Prince, Snakes, shapes, and gradient vector flow, IEEE Trans. Image Process, 1998, 7(3):

8 [16] Y Y Du, C H Zhou, Automatically Extracting Remote Sensing information for water bodies, Journal of Remote Sensing, 1998, 2(4): [17] HQ Xu, Modification of Normalized Difference Water Index (NDWI) to Enhance Open Water Features in Remotely Sensed Imagery, International Journal of Remote Sensing, 2006, 27(12): [18] JK Yu, Y S Huang, X Z Feng, Study on Water Bodies Extraction and Classification from SPOT Image, Journal of Remote Sensing, 2001, 5(3): [19] SG Chang, B Yu, and M Vetterli, Spatially adaptive wavelet thresholding with context modeling for image denoising, Proc. IEEE Int. Conf. on Image Proc, 2000, 9(9): [20] J.-L. Starck, E. J. Candès, and D. L. Donoho, The curvelet transform for image denoising, IEEE Trans. Image Process, 2002, 11 (6): [21] J.-L Starck, F Murtagh, E.J. Candes, D.L Donoho, Gray and color image contrast enhancement by the curvelet transform, IEEE Trans. Image Process, 2003, 12(6): [22] S Mallat, L Hwang, Singularity Detection and Processing with Wavelets, IEEE Transactions on Information Theory, 1992, 38(2):

Image denoising using curvelet transform: an approach for edge preservation

Journal of Scientific & Industrial Research Vol. 3469, January 00, pp. 34-38 J SCI IN RES VOL 69 JANUARY 00 Image denoising using curvelet transform: an approach for edge preservation Anil A Patil * and