Very High Resolution Satellite Images Filtering

Transcription

1 23 Eighth International Conference on Broadband, Wireless Computing, Communication and Applications Very High Resolution Satellite Images Filtering Assia Kourgli LTIR, Faculté d Electronique et d Informatique U.S.T.H.B., Bab-Ezzouar Alger, Algérie assiakourgli@gmail.com Youcef Oukil Département de Géographie E.N.S. de Bouzareah Alger, Algérie Y_oukil@yahoo.fr Abstract This paper introduces a new non linear filter suitable for artifacts (trees, cars, etc.) removing in Very High Resolution Satellite (VHRS) Images. It is based on the application of three filters of different sizes and shapes that permits each time to remove the artifacts smaller than the filter. These filters are based on the research of noisy pixels surrounded by homogenous area according to the three filters. First, tests have been conducted on Berkeley database images corrupted with different kinds of noise comparing the filter performances to that of anisotropic diffusion and bilateral filters. In a second stage, we applied our filtering technique to HRS images obtaining promising results. Keywords edge preserving smoothing filter; very high resolution satellite images. I. INTRODUCTION The very HSR (High Spatial Resolution) satellite data represent the surface of the Earth with more detail but this induces an increase of the internal variability within homogenous land cover units. In this context, image segmentation that permits to create regions instead of pixels was proposed as a pre-processing step [3], [4]. This is, then, used as input to a land-use classifier. In remote sensing, image segmentation is desired to provide meaningful object primitives for further feature recognition and thematic classification [7], [8], [], []. But, until now, there is no universal segmentation technique. Most of existing techniques suffer from the presence of geometric noise inherent to this kind of images. Indeed, the very high-spatial resolution of the image and the urban context induce a significant increase in geometric noise and artifacts such as vehicles, road markings, trees, occlusions or shadows that will disrupt the process of automatic extraction. To reduce the artifacts and make segmentation results more homogenous, we propose an iterative process of local filtering, which minimizes the average heterogeneity of the generated images by reducing artifacts. The first tests were carried out on the basis of images extracted from Berkeley database. We applied to these textured images different types of noise (Gaussian, Speckle and Salt & Pepper noises). Once validation is complete, we tested our filter on a very high spatial resolution satellite image taken from Google Earth software. The paper is organized as follows: In the next section, we present the proposed filter. The tests are presented and discussed in Section 3. The last section presents filtering results obtained on very high resolution satellite images. We end with a conclusion and perspectives for this work. II. COLOR IMAGE FILTERING Fragmentation is a key problem that exists in HSR image segmentation. It can cause undesirable holes in regions. Moreover, artifacts or small pieces of information can decrease segmentation efficiency. These problems exist more predominately in some solutions than in others and can cause difficulties for successful classification [5]. We referred to some classical filters usually employed with color images for smoothing images whereas preserving contours. The most successful filters are non linear ones such as anisotropic diffusion [6], bilateral filter [9], mean-shift filter [2]. However, none of them satisfy this goal perfectly: they each have exception cases in which smoothing may occur across hard edges and consequently affect the following processing steps (segmentation, classification, features extraction, etc.). Moreover, these filters are less efficient when the images are strongly textured and the denoising process may destroy the image edge structures. Our filter tries to overcome these problems by a new way to formulate smooth filtering. Usually, smoothing methods which are edge preserving are based on the general idea that the average is computed only from those points in the neighborhood which have similar properties to the processed point including the color of the processed point in the estimation. Our filter is not based on the comparison of the pixel under consideration to its neighbors, but is built around the idea of researching homogenous regions surrounding noisy pixels (artifacts in VHR satellite images). Once the homogenous regions are identified, their mean color is affected to the set of pixels they are surrounding. To find the homogenous regions, we use the three following filters: /3 $3. 23 IEEE DOI.9/BWCCA

2 F = F = 2 F 3 = Each filter will be used to scan the image looking for homogeneous areas (whose corresponding values equal ) surrounding noisy pixels. For this aim, the pixels whose value is are identified and the range (maximum minimum) of these pixels is calculated. If the estimated range value is below a threshold λ, the noisy pixels inside the uniform region are replaced by the mean value of the homogeneous ones. The process is iterated on each filtered image with the next filter (F, then F2 and finally F3). To assess I3SF filter performances, we compare it to a widely used non linear filters named anisotropic diffusion and bilateral filters. The former [6] uses local variation measures to iteratively average neighborhoods in order to avoid edge blurring. The latter [] [9] treats both space domain and range domain where weighted averaging is performed in the joint range-space domain. A brief explanation of both filters is given below. Perona and Malik [6] introduced the idea of anisotropic diffusion, that is, to smooth the image in the orthogonal direction to the gradient, U, and prevent, as much as possible, the diffusion across the direction of the edges (i.e. the direction of the gradient, U). They replaced the heat equation by the following one: (2) Where g( U ) is an edge-stopping function. g(.) has to be a nonnegative monotonically decreasing function with g() =. Different functions were used for g (.) giving perceptually similar results. The images in this paper were obtained using the second one: (3) The scale-spaces generated by these two functions are different: the first privileges high-contrast edges over lowcontrast ones, the second privileges wide regions over smaller ones. The behavior of the diffusion highly depends on the appropriate choice of the gradient thresholding scale parameter K. If a coarse-scale image with only the key edges is desired, a large K should be specified. And, to obtain a finescale image with all the detailed edges, a small K should be chosen. Bilateral filtering introduced by Tomasi and Manduchi [9] extends the concept of Gaussian smoothing by weighting the filter coefficients with their corresponding relative pixel intensities. Pixels that are very different in intensity from the central pixel are weighted less even though they may be in close proximity to the central pixel. This is effectively a convolution with a non-linear Gaussian filter, with weights based on pixel intensities. This is applied as two Gaussian filters at a localized pixel u with a neighborhood N(u), one in the spatial domain and one in the intensity domain (range) as: (4) The sum at the denominator is a normalization factor. Both functions: the closeness function (with a variance σ c 2 ) and the similarity function (with a variance σ s 2 ) are Gaussian functions of the Euclidean distance between their arguments. III. TESTS ON NOISY IMAGES The filtering methods have been applied to some textured images extracted from Berkeley database (Fig.). To better illustrate the effect of each filter and asses the performances of the proposed one; we focused on one image recorded in JPEG format. In our experiments, we tested the common types of noise: salt and pepper noise, Gaussian noise and speckle noise. The corresponding noisy images corresponding are shown in Fig..d, Fig. 2.d and Fig. 3.d The Gaussian white noise has a variance equal to.5, whereas the speckle noise is multiplicative obtained by adding to the image I a term n*i, where n is a uniformly distributed random noise with mean and variance.. Finally, the salt & pepper noise affects approximately 5% of pixels. Let us not that the noise is visible in homogeneous image regions but masked in textural regions. We, then, applied the three filters to these images with the aim to remove or at least reduce noise, strengthen the contours in order to improve their visibility whereas preserving textured areas contained in the image. For each filter, we estimated the PSNR, and evaluated the contour preservation. The results of applying anisotropic diffusion filter, bilateral filter, and our filter are given in Fig., Fig.2 and Fig. 3 for image 86. Different configurations were considered, each filter having its specific parameters. In general, For anisotropic diffusion filter, the number of iterations should not be to not too important since the filtering process is iterative inducing blur effect increases with the number of iterations, we have not gone further than three iterations. We tested several values (, 2, 3) for K threshold that controls the conduction and different values for the constant integration t (/5, /7, /5). The best results have been obtained considering the second g(.) function. The bilateral filter depends on fewer parameters. Its application requires the choice of window size (2w+) and parameters of spatial variance and intensity variance (σ c 2,σ s 2 ). 466

3 a) Gaussian noise it=3, t =/5, K=3 (PSNR=5,5) b) Gaussian noise w=3, σ c 2 =5, σ s 2 =. (PNRS=22,) c) Salt & Pepper noise λ=.8 (PSNR=4,97) d) Salt and Pepper %=,5 Fig.2. Filtering salt & pepper noise with a) anisotropic diffusion, b) bilateral, c) I3SF filters and d) noisy image. c) Gaussian noise λ=.85 d) Gaussian noise σ 2 =,5 (PNRS=7,46) Fig.. Filtering Gaussian noise with a) anisotropic diffusion, b) bilateral, c) I3SF filters and d) noisy image. a) speckle noise it=2, t =/7 K=2 (PSNR=7,69) b) speckle noise w=3, σ c 2 =5, σ s 2 =.2 (PSNR=9,59) a) Salt & Pepper noise i=t3, t =/7, K=3 (PSNR=5,89) b) salt & pepper noise w=5, σ c 2 =3, σ s 2 =. (PSNR=,79) c) speckle noise λ=.2 d) Speckle noise σ 2 =, (PSNR=2,) Fig. 3. Filtering speckle noise with a) anisotropic diffusion, b) bilateral, c) I3SF filters and d) noisy image. 467

4 The better results have been obtained using window sizes of and 7and σ c 2 = 3, σ s 2 =.. The last filter (I3SF) requires the adjustment of one parameter that is the threshold λ that controls the smoothing. The filtering results illustrated by the figures above show that: - The diffusion anisotropic filter, in all cases, is less efficient than the two others. Unfortunately, the smoothing of the image within the homogeneous areas is achieved through the blurring of color values which are containing the pertinent information. Indeed, the drawback of image denoising (smoothing) by anisotropic diffusion is that it tends to blur away the sharp boundaries in the image that help to distinguish between the larger-scale structures that one is trying to characterize. - With the bilateral filter, edges are better preserved, but texture tends to disappear. In fact smoothing of homogeneous zones is accompanied by a smoothing of the textured ones. Moreover, bilateral filter is ineffective with salt & pepper noise (Fig.2.b) The I3SF filter provides a better denoising when the image is corrupted with a speckle noise. It works well for image noise affected by a salt and pepper noise. There is no loss of contour and texture is preserved when λ is chosen less than.2. It seems able to restore blurred edges more effectively To summarize, the bilateral filter and I3SF filter are most successful in terms of filtering, edge enhancement and conservation of objects homogeneity. The anisotropic diffusion filtering induces a blurring of the image and loss of contours while for the bilateral filter a compromise must be made between smoothing and preservation of contours. Although, the I3SF filter requires more time execution, it seems to perform better for speckle and salt & pepper noises as noise reducing does not occur at the expense of contours preserving. IV. VHR SATELLITE IMAGE FILTERING Examples of the application of the proposed filter are now given for the filtering of very high-resolution satellite image taken from a densely built-up area extracted from Google Earth-software. Fig. 4.a shows the image extracted and Fig.4.b. shows the result of the application of the mean-shift segmentation and we can observe that the man-made objects are not well delineated because of cars, trees and objects on roofs. Therefore, we applied the three filters obtaining the Fig. 4.c, Fig. 4.d and Fig. 4.e. Again, different configurations were considered. In general, the anisotropic diffusion filter produces too much smoothing that remains even when we reduce the number of iterations and increase the integration constant. The bilateral filter better preserves the contours for a suitable choice of window size and the two variances values. Moreover, the second is high, the stronger the smoothing and worse will be the delineation of objects on the ground. The last filter (I3SF) provides images were homogeneous areas are smoothed and some artifacts are eliminated (objects on the roof) without affecting contours (see area zoomed in Fig. 5). a) Original image b) Segmented image using mean-shift algorithm c) Bilateral filter w=5, σ c 2 =3, σ S 2 =.2 468

5 a) Original image d) Anisotropic diffusion filter K=2, Δt=.5, iter = 4 b) Bilateral filter w=5, σ c 2 =3, σ S 2 =.2 Fig.4. VHR satellite image filtering. e) I3SF filter λ=.8 In Fig. 6, we give the range images corresponding to the four areas zoomed of Fig. 5. Without filtering (Fig.6.a), edge detection using range estimation can t well distinguish important and unimportant edges. Smoothing provides a partial solution to this issue. In fact, by introducing smoothing (Fig.6.b and Fig.6.c), we keep edges with strong gradient and discard edges with weak gradient. But due to filtering on image, results in more blurring. Hence, smaller edges are not preserved and the actual edge location is shifted due to blurring operation. Using the filter proposed (Fig. 6.d); the detection and handling of the edge structure of the input image are better achieved since main structures such as buildings are well highlighted (strong edges). c) Anisotropic diffusion filter K=2, Δt=.5, iter = 4 Fig.5. Zoom on filtered images. d) I3SF filter λ=.8 469

6 a) Original image V. CONCLUSION The goal of the nonlinear smoothing is to improve the accuracy of the segmentation by preserving significant changes in image intensity, whereas smoothing random noise fluctuations. It is usually employed as a pre-processing step for segmentation purpose yielding to a better result. The results obtained show that I3SF filter is able to effectively denoise the image without blurring edges as noise reducing does not occur at the expense of contours preserving. In experiments, it demonstrated good performance on VHR satellite images resulting in less edge blurring and artifacts reduction. REFERENCES b) Bilateral filter w=5, σ c 2 =3, σ S 2 =.2 c) Anisotropic diffusion filter K=2, Δt=.5, iter = 4 [] D. Barash, A Fundamental relationship between bilateral filtering, adaptive smoothing and the nonlinear diffusion equation, IEEE Trans. on Pattern Analysis and Machine Intelligence, vol.24, No.6, pp , 22. [2] D. Comaniciu, P. Meer. Mean Shift: A Robust Approach toward Feature Space Analysis, IEEE Trans. pattern Analysis and Machine Intelligence, vol. 24, pp , (22). [3] A. Guarnieri, A. Vettore. Automated Techniques for Satellite Image Segmentation ISPRS Commission IV, Symposium 22, Volume XXXIV Part 4, 22 : Geospatial Theory, Processing and Applications. Editor(s): Costas Armenakis, Y.C.Lee. July 9-2, 22, Ottawa, Canada. [4] X. Hu, C.V., Tao, B. Prenzel. Automatic segmentation of highresolution satellite imagery by integrating texture, intensity and color features, Photogrammetric Engineering and Remote Sensing, vol.7: pp , (25). [5] X. Jie, S. Peng-fei, "Natural color image segmentation", in Proceedings of International Conference on Image Processing ICIP, pp , 23. [6] P. Perona, and J. Malik, Scale-space and edge detection using anisotropic diffusion, IEEE Trans. Pattern Analysis and Machine Intelligence, vol.2,pp , 99. [7] M. Singha, K. Hemachandran. Color Image Segmentation for Satellite Images, International Journal on Computer Sc ience and Engineering (IJCSE), vol. 3 (2), pp , 2. [8] J. Tian, D. Chen. Optimization in multi-scale segmentation of highresolution images for artificial feature recognition, Int. Journal of Remote Sensing, vol. 28 (2), pp , (27). [9] C. Tomasi, and R. Manduchi, Bilateral filtering for gray and color images, Proceedings of the IEEE International Conference on Computer Vision (ICCV), Bombay, India, pp , 998 [] B. Varga, K. Karacs. High-resolution image segmentation using fully parallel mean shift, EURASIP Journal on Advances in Signal Processing 2. [] C. Zhong, Z. Zhongmin, Y. DongMei, C. Renxi. Multi-scale segmentation of the high resolution remote sensing image. In: Proceedings of IEEE International Geoscience and Remote Sensing Symposium (IGARSS 25), pp ,(25 d) I3SF filter λ=.8 Fig.6. Zoom on range of filtered images. 47