A Structured Approach to Automated Crater Detection

Introduction: A structured methodology aimed at the automated recognition of impact craters on planetary surfaces is presented. The initial phase focus on edge detection; this is followed by a crucial step, in which a template matching procedure is employed to create a probability volume from which the best candidates are selected and undergo a third phase, designed to detect the centers and estimate the dimensions (radius) of craters which are then plotted on the images. Methodology: The approach we present proceeds according to a similar scheme, used by other workers in this field [1-5]. It comprises three main phases, which are shortly described below (for a fuller description, see [6, 7]), and illustrated in figure 1 through their sequential application to a MOC/MGS wide angle image (figure 1 (a)). Edge detection. In this phase the objective is to identify, in the scene image, regions that correspond to crater rims. To achieve this goal we use an edge detection operator which incorporates local information and thus constitutes an improvement relative to classic edge detectors. The result is a binary image that will serve as input for the next phase of the method (figure 1 (b)). Template matching. This process involves crosscorrelating a template with the scene image and computing a measure of similarity between them. The template employed in this work is a simple black and white circular crater model (see figure 1), suited to the binary image that resulted from the edge detection. The Fast Fourier Transform, a proven method applied in the frequency domain, was used for the actual computation of the correlation between the scene image and the series of templates with different sizes. The values (one per pixel and per template dimension) are normalized and collected into a probability volume (figure 1 (c)), a stack of r planes (r being the range of values used for template radius) each containing u x v pixels (the size of the image). Crater recognition. Assuming that a crater produces an identifiable signature on the probability volume, we look for all the local probability maxima and sort them out in a number of steps involving the analysis of their neighborhoods according to morphological features (dimension and roundness are considered). The end result of this cyclic process (which runs as many times as there are planes in the probability volume) is the elimination of weaker candidates and the identification of the probable centers of craters, along with the corresponding radius (figure 1 (d)).