A Contribution to Image Registration in Satellite Imaging. M. Tehami, N. Taleb

Transcription

1 A Contribution to Image Registration in Satellite Imaging. Tehami, N. Taleb laboratoire Telecommunications and Digital Signal Processing Laboratory, Département d électronique faculté des siences de l ingénieur, Université Djillali liabès,2200 Sidi Bel-Abbes, Algeria E- mail : t_malika2001@yahoo.fr Tel, Fa: Abstract With the increase in the number of images collected every day from different sensors, automated registration of multi-sensor/multi-spectral images has become an important issue in remote sensing. In the near future, image registration will be one of the basic image processing operations in remote sensing. These systems will provide large amounts of data representing multitime or simultaneous observations of the same features by different sensors. The objective of this paper is to present an automatic registration algorithm. The designed algorithm has the capability to register images, even when applied to remote sensing images (multidate, multispectral, and multisensor satellite images). This algorithm, introducing local processing on areas of interest consisting of piels belonging to image edges only, is fast compared to all traditional algorithms where entire images are searched. The algorithm was tested on SPOT and T images from Agricultural,Desert, Amazon areas. In all cases the obtained results are very encouraging. Keywords: image registration; image processing; remote sensing; matching. 1. Introduction Image registration is the process of matching two images so that corresponding coordinate points in the two images correspond to the same physical region of the scene being imaged. It is a classical problem in several image processing applications [1] where it is necessary to match two or more images of the same scene [2]. Due to increasing amount and diversity of remotely sensed data, with the dramatic increase in data volumes and types of sensors [3], image registration is becoming one of the most important issues in remote sensing. An important consideration for selecting the registration method to be employed for a given problem is the source of misregistration [4].The source of misregistration is the cause of the misalignment between images. It may be due to change in the sensor position, viewpoint and viewing characteristics or to object movement and deformation. The motivation of this study is a new algorithm for automatic satellite image registration, this algorithm introducing local processing on areas of interest consisting of piels belonging to image edges only, very fast compared to all traditional algorithms where entire image are searched [5]. A description of the mentioned algorithm and presentation of some satellite image registration results are the objectives of this paper.the techniques process the piels belonging to image edges only, instead of processing all image piels, the correlation coefficient is used as a similar measure and only the piel who has the best match from all piels in the image leads to the best global registration parameters. The paper is organized in the following way. The different steps of the general method for global and local image registration are presented. Section 2 gives a technique of selecting control points. The registration parameters determination is presented in Section3, Section 4 shows the preliminary eperimental results of registering images taken from Landsat and Spot satellites and Section 5 summarizes the work. 2. Control points selection The correspondence between two images was computed only for a selected set of so called control points. The control points can be chosen manually by selection of a region of interest and/or can be taken to lie on a regular grid. It has to be noted that the change and misalignment only appear in those regions where strong object edges are present in the these images. Clearly misalignment between images can only occur at the image edges [6], and therefore one easy way to speed up the algorithm and to reduce the computation time is to process piels belonging to image edges only instead of processing all image piels. Obviously, the less comple the image, the faster the algorithm. The first step in the selection of the control points consists of finding image edges. The original (input) image is first sharpened using highemphasis filtering [7]. The gradient image and the edge-direction map are then computed using the Sobel operator. Combining the gradient image with the edge-direction map and the threshold map, and applying a local connectivity algorithm [8] will

2 permit the derivation of the better edge map image (binary image). The following is a summary of the different steps to be followed in the selection of control points. 1. Choose a moving window W of w w piels that will sweep over all the piels in the edge map image. The size of this window depends on the size of the images themselves. This window will be used later for the measure of match (according to a certain similarity criterion). 2. ove the above window, line by line, through the edge map image. If a piel value is 1 (piel belonging to an edge), then this piel is considered to be a control point. 3. To avoid taking all piels belonging to an edge inside the same window as control points, take just one piel as a control point and assign a 0 value to all the rest of the piels in the same window. 4. Repeat steps 2 and 3 until going through all piels belonging to all edges. 3.Registration Parameters Determination The registration parameters are determined by carrying out the following steps. 1.Choose a moving window that will sweep over all the piels in the edge map image, the size of this window depends on the size of the images themselves, some typical values are a (30 30) window for a ( ) image, and a ( 20 20) window for a ( ) image (Fig1.a). 2. ove the above window, line by line, through the edge map image. 3. If a piel is 1 (piel belonging to an edge) (Fig1.b) then: 4. At the same piel position on the original image take piels belonging to a window of the same size centered at that same position (Fig1.c). 4.1 Assuming that for ( ) images misregistration parameters would not eceed a ±12 degree angle for image rotation and a ±70 piel shift for image translation (reasonable practical assumptions), ±8 degree angle and ±30 piel shift for ( ) images, rotate this window by an angle of 12 degrees (Fig1.d) [8] and correlate it with the corresponding non- rotated window in the reference image taking into consideration the amount of the maimum translational shift (Fig1.e). Save the translational shift giving the largest correlation coefficient. This coefficient should not be smaller than Repeat step 4.1 after each rotation of a +1 degree angle of the preceding window until reaching a rotation angle of +12 degrees. After each rotation keep saving the angle of rotation and the translational shift that result in the largest correlation coefficient at all times. At the end these two values correspond to the eact local registration parameters. 4.3 Repeat step 4.2 until going though all piels belonging to all edges. 4.4 At the end of step 4.3 locate piel where a maimum correlation coefficient has been obtained, this coefficient should not be smaller than 0.9.At this location, rotate and translate the entire original image using the registration parameters obtained in steps (4-3), the result is the registered image (Fig1.f). All these steps are recapitulated in the flowchart diagram of figure (2) for a ( ) image. 4. Eperimental Results In this section, the performance of the proposed registration technique in remote sensing image sequences is evaluated eperimentally. In order to test the algorithm and demonstrate its feasibility for different types of images, preliminary results are presented in this section. Images from Landsat T and Spot Satellites have been used in the eperiments. These data sets used here consisted of sequences of ( ) and ( ) images. Each image has a gray value resolution of 8 bits,i.e. 256 gray levels. The data sets are multidate, multispectral and multisensor remote sensing image, acquired on a R imaging system operating in the department of electrical engineering and computer science at George Washington University. The results of the application of the proposed registration technique to multidate, multispectral, and multisensor remote sensing images are shown in figures (3) through (6),respectively. In these figures, four images are chosen for each data set : (a ) the original (sensed, input) image, (b) the reference image, (c) the edge image obtained from one of the original image, (d) the registration of the two images. Figure (3) shows the registration of two images taken from Landsat T sensor, band 5 on different dates 09/10/91 (T945) and 07/10/90 (T905). They correspond to an agricultural region near Itapeva, Sao Paulo. The image (T945) was taken as the original one and (T905) as the reference one. Figure (4) shows the registration of Desert region images taken from Landsat T sensor band 2, in different dates,06/07/92 (T922) as the original one and 07/09/95 (T952) as the reference one. Figure (5) shows the registration of two agricultural images of year 88 at different sensors, (Spot Sensor, SP885) as the sensed image and (Landsat T,T885) as the reference one. Figure (6) shows the registration of two Desert images taken from Landsat T sensor of year 88 on different bands, (band4,t884) as the sensed image and (band7,t887) as the reference one.

3 The parameters of the transformation and the edge thresholds for the test images shown in Figures 3-6 are listed in Table I, Where he unit for ' and 'y is in piels and for T in degrees. 5. Conclusion An automated algorithm for the registration of satellite images has been described and a new approach to the registration of remote sensing image is proposed. The method involves the etraction of a set of control points by using an edge detection approach in the sensed (input, original) images because misalignment between similar images can only occur at image edges. The registration parameters of every control point are determined by means of template matching, using the correlation based similarity measure. oreover, the proposed algorithm demonstrates the capability to register two images and detect the different misalignment in similar images. This algorithm is fast compared to all traditional algorithms. Also the proposed approach yields to better global registrations,when applied to remote sensing images, but also in any other registration problem. [8] R. Bernstein, Digital Image Processing of Earth Observation Sensor Data, IB. J Research and Development 20, January 1976, pp Images T945 T905 T922 T952 SP885 T885 T884 T887 Table I Edge ' 'y 'T thresholds Correlation Coefficient % % % % 0.99 References [1] G. Câmara. R. C.. Souza, U.. Freitas, and J. Garrido. Spring: Integrating remote sensing and gis by object oriented data modelling. Computers&Graphics,20(3): ,ay- June [2] J. E. Canny. A computational approach to edge detection. IEEE Trans. On Pattern anal. And achine Intell., 8(6) : , Now [3] H. Li, B, S. anjunath, and S. K. itra. contour-based approach to multisensor image registration. IEE Transactions on image Processing,37(12): ,Dec [4] L.. G. Fonseca and B. S. anjunath. Registration techniques for multisensor remotely sensed imagery. Photogrammetric Engineering and RemoteSensing,62(9): ,Sep [5] P. Aleander, Array processors in medical imaging. IEEE computer,17-30, June [6] A. Rosenfeld, A.C. Kak, Digital picture processing (second edition, volume1),pp ,1982,academic Press. [7] N. Taleb, L. Jetto, Image registration for application in digital subtraction angiography, Control Engineering Practice, vol.6,pp, , 1998.

4 oving window ( ) Piel belonging To edges Edge map Image E (a) (b) Edge map Image E E(I,j)=1 at position (i, j ) original Image O O(i,j) O(i,j) At position At (i,,j) position (i,j) Regular grid 12 Rotated piel Interpolated Piel value Window from Original image Rotated by a (-12 ) angle (c) (i, j) (d) Search area (N N) (non rotated window in reference image). Window from original Image where piels were rotated by a (-12 ) angle. N (i,j) N 'T ' 'y Original image Register image Reference image (e) (f) Figure (1): Schemes of general registration method

5 Start Image Sharpening of Original image Acquisition of gradient Image and Edge Direction ap Derivation of Edge ap B(I,J) by applying Local Connectivity I =1, J =1 Net Piel B(I, J) No No Are All Piels in B(I,J) Generated? B(I, J)=1? Yes Rotate Window Centered at (I,J) on the Original Image From 12 Deg to +12 and Correlate it with corresponding window on reference image Considering the aimum Shift. Save Angle of Rotation and Translational Shift Giving a aimum Correlation Yes Are all Piels in B(I,J) Generated? No Net Piel B(I,J) Locate piel where a maimum correlation coefficient has been obtained At this location, rotate and translate the entire original image using the registration parameters obtained. end Figure (2): Flowchart diagram of the general registration method

6 (a) (b) (c) (d) Figure (3): Registration of two Agricultural images taken from Landsat (T) sensor band 5 at different times: (a) The original image (T 945). (b) The reference image (T 905). (c) The edge image obtained from the original image. (d) Registration of two images. (a) (b) (c) (d) Figure (4): Registration of two Desert images taken from Landsat (T) sensor, band 2 at different times: (a) The original image (T922). (b) The reference image (T952). (c) The edge image obtained from the original image. (d) Registration of two images. (a) (b) (c) (d) Figure (5): Registration of two agricultural images of year 88 at different sensors, band 5: (a) The original image (Spot sensor, SP885). (b) The reference image ( Landsat, T885 ).(c) The edge image obtained from the original image. (d) Registration of two images. (a) (b) (c) (d) Figure (6): Registration of two Desert images taken from Landsat T sensor, of year 88 at different bands: (a) The original image (band 4, T884). (b) The reference image (band 7, T887).(c) The edge image obtained from the original image. (d) Registration of two images.