Showing papers on "Image registration published in 1976"

Experimental examination of similarity measures and preprocessing methods used for image registration

Abstract: The criterion used to measure the similarity between images and thus find the position where the images are registered is examined. The three similarity measures considered are the correlation coefficient, the sum of the absolute differences, and the correlation function. Three basic types of preprocessing are then discussed: taking the magnitude of the gradient of the images, thresholding the images at their medians, and thresholding the magnitude of the gradient of the images at an arbitrary level to be determined experimentally. These multitemporal registration techniques are applied to remote imagery of agricultural areas.

Image Registration Error Variance as a Measure of Overlay Quality

Abstract: When one image (the sigal) is to be registered with a second image (the signal plus noise) of the same scene, one would like to know the accuracy possible for this registration. This paper derives an estimate of the variance of the registration error that can be expected via two approaches. The solution in each instance is found to be a function of the effective bandwidth of the signal and the noise, and the signal-to-noise ratio. Application of these results to LANDSAT-1 data indicates that for most cases registration variances will be significantly less than the diameter of one picture element.

Automatic Threshold Setting for the Sequential Similarity Detection Algorithm

Abstract: This correspondence presents a method for automatic setting of both constant and increasing thresholds to be used with the sequential similarity detection algorithm (SSDA) for a fast registration of digitized images. No a priori knowledge of image statistics is required. The usefulness of the method is proven in the cloud tracking.

Evaluation of Landsat image registration accuracy

Abstract: The Large Area Crop Inventory Experiment (LACIE) is an attempt to demonstrate the capability to forecast the annual production of major crops such as wheat and corn. Good image registration of data acquired on different dates is one of the key assumptions made in LACIE. This paper describes an algorithm to measure the accuracy of the current registration procedure. This algorithm employs a modified version of sequential similarity detection algorithm (SSDA). Based on over 264 registration checks, it was found that the root-mean-square of registration errors was 1.0 pixel. The failure rate of our registration checking algorithm was less than 10 per cent and the standard deviation of the accuracy of this algorithm was less than 0.2 picture element.

Color television camera with internal registration

Abstract: A color television camera including a plurality of camera tubes and an image registration system in which information is projected, via the beam path of the viewer, to a beam splitting optical system. A registration image is projected onto each of the camera tubes by providing a suitable end face of one of the prisms of the splitting optical system with a reflective layer or by arranging a mirror or a prism at this area.

Invariant Features for Quantitative Scene Analysis.

Abstract: : This report summarizes a study of map matching technique for precision guided reentry vehicles using invariant measurements and scene analysis techniques. A particular case of matching optical and side looking radar images is considered in detail. Both geometric and sensor transformations are developed. Edge registration including a new vector representation of scenes which may be used to derive edge operators and unifies most edge operators currently in use, is presented. Invariant measurements depend on algebraic invariant theory and perceptual experience. The set of moment invariants are considered which are invariant to translation, rotation, size and mirror images, for computer for use. Finally, hierarchical search techniques which are logarithmetically efficient are presented for map matching. These techniques are illustrated with correlation measures of picture values, edge measurement and moment invariants. These techniques appear very effective for map matching. (Author)

Analytical and experimental design and analysis of an optimal processor for image registration

Abstract: The author has identified the following significant results. A quantitative measure of the registration processor accuracy in terms of the variance of the registration error was derived. With the appropriate assumptions, the variance was shown to be inversely proportional to the square of the effective bandwidth times the signal to noise ratio. The final expressions were presented to emphasize both the form and simplicity of their representation. In the situation where relative spatial distortions exist between images to be registered, expressions were derived for estimating the loss in output signal to noise ratio due to these spatial distortions. These results are in terms of a reduction factor.

Geometric assessment of image quality using digital image registration techniques

Abstract: Image registration techniques were developed to perform a geometric quality assessment of multispectral and multitemporal image pairs. Based upon LANDSAT tapes, accuracies to a small fraction of a pixel were demonstrated. Because it is insensitive to the choice of registration areas, the technique is well suited to performance in an automatic system. It may be implemented at megapixel-per-second rates using a commercial minicomputer in combination with a special purpose digital preprocessor.

Evaluation of Image Registration Accuracy in the Large Area Crop Inventory Experiment

Abstract: The Large Area Crop Inventory Experiment (LACIE) was created to demonstrate the capability of forecasting the wide area annual production of major crops, such as wheat and corn, from multispectral LANDSAT imagery and meteorological data. The current classification algorithm utilizes multi-temporal imagery from different biological phases which are spatially aligned or registered. It is of paramount importanceto investigate and evaluate the accuracy of the registration process, because the number of pixels and the area classified are directly affected. A new technique ha~ been developed at the NASA Johnson Space Center to evaluate image registration accuracy. The first step in the registration of an image with a previously acquired reference image is to extract a 234 scan line by 354 pixel area from a full frame LANDSAT image based upon the best estimate of the LANDSAT position. The second step is to find a best match position by correlating two images on a digital computer. This is being performed at NASA Goddard Space Flight Center, and the current scheme uses binary edge images derived from the original gray level images for correlation. Our task was to evaluate the registration accuracy in a set of LACIE image pairs which were registered with the above process. It was found that they are registered very closely and it suffices to search for a best match within a small range, say plus and minus five pixels. The approach used involves two steps: the first is to register small subimage pairs extracted from the two LACIE images. For each subimage pair a best match is determined solely by shifting along the x and y coordinates because the dominant transform for small subimages (say 27 by 27) is the x and/or y shift only. The second 2B-2 step is to calculate a six parameter linear transform from the (x, y) coordinate system to a new coordinate system (p, q) using the local shift pairs (8X., 8Y.). The six coefficients can be d~term!ned by the least square criteria. The LACIE images are small and notation and scale changes cannot be estimated reliably thus, only the shifts 8X and 8y are computed. With this approach there is need for a technique which will ~ rapidly correlate two small subimages. It was found that methods using edges did not perform well on the small subimages being used. We chose instead to work directly with gray level images rather than edge images by employing a modified sequential similarity detection algorithm. Experiments were ca~ried out for 44 LACIE sample segments, each of which had four acquisitions. By using channel 2 images registration accuracy was evaluated on all the six possible combinations per sample segment. It was observed that the range of 8X and 8y are restricted within plus and minus 2 pixels and 1 pixel, respectively. One reason for the observation that the distribution of 8X is wider than that of 8Y is that the LANDSAT images are less sharp and edges are less reliable along the x axis. The root mean square of registration error is .99 pixels which satisfies the accuracy objective of the current registration procedure.

Architecture and data processing alternatives for the tse computer. Volume 4: Image rotation using tse operations

Abstract: The tse computer's capability of achieving image congruence between temporal and multiple images with misregistration due to rotational differences is reported. The coordinate transformations are obtained and a general algorithms is devised to perform image rotation using tse operations very efficiently. The details of this algorithm as well as its theoretical implications are presented. Step by step procedures of image registration are described in detail. Numerous examples are also employed to demonstrate the correctness and the effectiveness of the algorithms and conclusions and recommendations are made.