PARAMETER SPACE TECHNIQUES FOR TEMPORAL IMAGE REGISTRATION (DIGITAL, ANALYSIS)

PAUL EDWARD ANUTA, Purdue University

Abstract

The purpose of this research is to investigate the use of parameter space techniques for achieving registration between images which have undergone significant change due to temporal effects of the environment. The parameter space method investigated in this thesis uses the Hough transform to create a new type of image which is used to find intersections of lines to be used as control points for registration. This method can identify linear intersections regardless of rotation and other low order distortions. Two other methods for achieving image registration are studied. The first is conventional block correlation of image pairs to determine translational shift. The second is block correlation of edge images derived from the original images. This method is an attempt to utilize invariant edge information rather than the highly variable scene content itself and also permits computational saving since a binary image is correlated. The three methods are evaluated analytically but comparison on a standard basis became difficult due to the nonlinearity of the methods. A random image simulation algorithm was implemented which allowed evaluation of all three methods under the same conditions. Simulator output is generated for 9 sets of parameters describing 3 edge widths and 3 characteristics of temporal change noise. For each case ten temporal signal-to-noise ratios are evaluated. The results indicate that the parameter space technique works well over a wide SNR range. Variances are lower for image correlation and much higher for edge correlation for SNR below 3 or 4. Results were also obtained for the parameter space method applied to Landsat data. One acquisition of high resolution (30 m) thematic mapper was available and an error radius standard deviation of .92 pixels was achieved. A temporal sequence of five lower resolution (80 m) Landsat multispectral scanner (MSS) scenes was available and four of the five cases produced accurate results for both the parameter space method and image correlation. The fifth was temporally distant and satisfactory correlation was not possible between this scene and any of the others. The parameter space technique also had difficulty in achieving satisfactory results but through the use of a scene heuristic interpretation of the correct control point was possible. The method is considered to be a promising candidate for temporal registration tasks in an environment of geometric distortion other than simple translation.

Degree

Ph.D.

Subject Area

Electrical engineering

COinS