GISdevelopment.net ---> AARS ---> ACRS 1990 ---> Poster Session Q

The registration and mosaic of digital images remotely sensed

Yang Venjiu
Center for Remote Sensing in Geology,
Minsitry of Geology and Mineral Resources,
29 Institute Road, Beijing, China


Abstract
Three available registration ways of digital images and three methods of control point selection which include: (1) automatic selection of control point between on-line images by using correlation technique, (2) interactive selection of control point between on-line image and off-line map/image, (3) calculation of the coordinate of control point. The digital mosaic processing is carried out based on the registration. The techniques of gray level adjusting; optimum juncture point selecting and gray level smoothing are adopted in the processing. On the mosaicked image the seam is well eliminated. The high quality mosaic image can meet the requirement of the application and research of remote sensing information. A set of computer software for these techniques was developed.

Introduction
The registration and mosaic of digital image are interrelated and independent to each other in processing of remote sensing information. For registration generally the reference image and the one to be corrected are input to the digital image processing system and interactively selecting control point on both images on the system monitor. Based on the control points the spatial transformation model is fitted. The image to be corrected is processed according to the model and the result image that matches to the reference one geometrically is generated. However, as the application of remote sensing technique is getting more and more widespread and profound, but it has still not satisfied the requirement of application and research if only this registration way is used. In this paper, three different registration ways of digital image are presented. After spatial transformation model is defined registration. Therefore, the different methods of control point selection are used for each registration was.

Mosaic of digital images is based on registration of the images. However, since the image to be mosaicked is not acquired simultaneously for a region, the gray levels or hues usually present some charges between the images. It brings about a spurious artificial seam on the mosaic. A digital mosaic technique, which may produce high quality mosaic, is presented in this paper.

Registration of digital images
  1. Registration of On-Line Images.

    This is a common way of registration. The control point pairs of the reference image and the one to be corrected are selected by user on the monitor. Based on the control point pairs the spatial transformation model is fitted. It causes the output image to be congruent with the reference one geometrically.

    After spatial transformation model is defined the corrective selection of control point pairs is the only way to improve the registration accuracy. Visual selection is a general and safe method, but it is time consuming and no high accuracy of location. In order to selecting with high accuracy and save time, the automatic selection techniques of control point are adopted in some conditions. The determination of same ground point from images acquired in different time actually is a problem of pattern reorganization. On image, any surface target is a spectral response of surface characteristics. In most cases the same targets are similar or consistent in spectral features. In other words their variation pattern of spectral features relative to the surrounding object are similar or consistent. Therefore, the location of same ground point for different, temporal images is a pattern-matching problem.

    The correlation technique is used to locate the same ground points. For a local neighborhood of two images their correlation measurement is defined by following equation:


    where f1 is the image to be corrected and f2 is the reference image. (I,j) are indices in a I*J pixel window area W. The window W is located within a M*N pixel searching area S. F1 is the mean of gray level of all pixels in the window W on image is the mean of gray level of all pixles in the area with same dimension as the window W within the searching area S on image f2.

    User is asked to select a point on the image f1 and the window W is defined, as a neighborhood located with the point as the center. Then the user selects a point arbitrarily near the corresponding position on the image f2 so the searching area S is delimited, which is an area with the point as the center. According to the equation the area S is searched correlatively by the window W. The maximum value of R (m,n) is found out and (m,n) are its coordinates. So that this point and the one selected by the user on the image f1 are as a control point pair. To do this repeatly until enough pairs have been selected to fit the transformation model for image registration. 2

  2. Registration between On-Line Image and Off-Line Map/Image

    The registration between on-line image and off-line map/image is the problem encountered frequently in application of remote sensing information. In an ordinary way, first the off-line map/image is digitized into an on-line image, then registration is conducted according to the registration way of on-line images, This process is tedious and restricted by hardware. A new registration way in which only the image to be corrected is needed to input to the system

    Tem and the reference map/image is used by off-line, User can select a control point on the system monitor and find the corresponding position on the off-line map/image. The coordinates of the position are typed in the system at terminal and converted to the coordinates as consistent with the image to be corrected. So that a control point pair of on-line image and off-line map/image is determinate. The user repeated selecting control point on both on-line and off-line ones until the control are as such as enough. Based on these control points the on-line image is registered to the off-line map/image geometrically.

  3. Generating CNPT File by Typing the Coordinates of Images into the System to Achive Image Registration

    Sometimes image registration is not completed by using the methods of control point selecting as mentioned above. In this situation the coordinates of control point need to converted and calculated in the light of specific conditions. Before spatial transformation of image new coordinates of control point are typed in at terminal by user and a CNPT (control point) file is generated. The image registration is brought about based on the control point pairs in the file.
Mosaic of digital images
Then two or more overlapping images are combined for mosaic; there would be no exat mosaic geometrically without accurate registration. Nevertheless, it must be pointed out that if one wants to acquire a high quality mosaic not only an accurate registration but also an appropriate matching of gray level or hue for the images to be mosaicked is needed. It is important to make good choice for the candidates of images to be mosaicked. Moreover, the preprocessing, which is designed to improve, geometric and radiometric qualities of the images have to be conducted before putting them together. The specific techniques of image mosaic adopted as follows.
  1. Adjustment of Gray Level.

    For some reasons, the difference of gray levels or hues between selected images is in existence frequently. The gray level matching between the images must be conducted 4. The method is taking one image as a reference and transforming the other to match the reference one in histogram. The mathematic formula used in histogram matching is:


    Fig. 1 A sketch map of mosaic and juncture point selection.

    Where IN is the gray level transformed. IA and IB are gray levels of the reference image and one to be adjusted respectively, sa and sb are variances corresponding to IA and IB, and are means of IA and IB respectively.

  2. Selection of Juncture Point of Mosaic

    A slide window is adopted to search the position where the difference of gray levels is the minimum in a neighborhood in overlapping area of two adjacent images line by line in order to eliminate the seam, which is often apparent in mosaic. The algorithm utilized is of the following form:


    Where IA(N+i) and IB (n+i) are gray levels of the overlapping area of two adjacent images respectively. Let the window size W be much less than the width of the overlapping area. The value of n which causes D(n) to get the minimum is found out by using the window and as a juncture point of a line. To do this repeatly, untill all juncture points are searched for all of the lines. Sometimes the seams are apparent along the direction perpendicular to the mosaic direction if the images are mosaicked directly by the juncture points. (Due to two juncture point of two adjacent lines may be far from each other). To avoid this problem, except the first line, the candidate juncture point will be selected from the following three points. One is the point in the same column. The other two points are situated on the left and the right of this point. Then the juncture point is selected among the three points with the slide window. Fig. 1 is a sketch map of mosaic and juncture point selection.


    For a color image mosaic the juncture line of the first band of the composite image is as a juncture line for all three bands in order to avoid artificial hues due to each band using different juncture line.

  3. Gray Level Smoothing of Neighborhood of Juncture Point

    Sometime there is still the difference of gray levels of both sides of the juncture point. The smoothing processing technique is adopted to eliminate the difference. The algorithm used is as follows:


    Where INi is gray level smoothed, IAi and IBi are gray levels of the juncture point. The smoothing processing. K is a weight. W is much less than the width of the overlapping area and its size is specified as a parameter.
The result of mosaic processing and evaluation
Fig. 2 is a false color mosaic image from Landsat TM image (the left part of the mosaic) and MSS image (the right part of the mosaic) of a region in Yunnan, China. First the MSS image was warped to congruent with the TM image geometrically by the registration way between on-line images. Then, the mosaic technique is used for both images to achieve the final result image.

Four scenes of Landsat MSS false color image of Xinjiang, China have been mosaicked by using the technique of the registration and mosaic of images presented in this peper. Fig.3 shows a portion of the mosaic image which includes the overlapping area of the four scenes. There is not any seam on the mosaic. A number of points of the overlapping area were selected for both original images and the result image of mosaic to count the error of registration and mosaic. The mean square error is only 0.75 pixel. The images achieved by the techniques presented in this paper show high qualities in both geometry and radiometry.


Fig. 2 A mosaic image form Landsat TM and MSS images. On the image the left part is TM data and the right part is MSS data.


Fig.3 The mosaic image of the overlapping area of four scanes of Landsat MSS false color images.

The results shown the techniques of registration and mosaic available and effective in digital image processing.
    The software development of image registration and mosaic
    Five functions were developed on 12S SYSTEM 101 in order to put the techniques presented to realize. The new functions are: (1) Automaticly selecting control point between on-line images and generating a CNPT file, (2) interactively selecting control point between on-line image and off-line map/image and generating a CNPT file, (3) typing control point data into based on the variances and means of two images and (5) juncture point and put the images to be mosaicked together.

    Conclusion
    A complete set of methods of digital images registration and mosaic was established and the software for these techniques were developed by the author in integration of on-line images the speed and accuracy of control point selecting are increased. The problem of the registration between on-line image and off-line map/image is solved. The digital mosaic processing technique of gray level adjusting, juncture point selecting and gray level smoothing is adopted successfully and the high quality mosaic image can meet the requirement of the application and research of remote sensing information.

    References
    1. System 101 image processing operating system user manual, version 3.1, International Imaging System,(1981), 134.

    2. W. K. Pratt, Digital Image Processing, Wiley Interscience, New York, (1978), 344.

    3. Floyd F. Sabins Jr., Remote Sensing Principles and Interpretation, W. H.Freeman and Company, San Francisco, (1978), 79

    4. Shigekatsu Horri et al., Digital Mosaic Processing, Proceedings of the Eighteeth International Symposium on Remote Sensing of Environment, (1984), 1785.