Two-stage Contour Detection Method for Small-scale Disk-resolved Celestial Datasets

In the astrometry of the disk resolved objects, contour detection acts as an important step, and it is equivalent to a binary classification process of classifying pixel points as contour points and non-contour points. For the purpose of automated extraction of hierarchical contour features, enhancement of contour detection precision, and accommodation of either limited-scale disk-resolved celestial datasets or preliminary-phase observational campaigns, a two-stage contour detection method for small-scale disk-resolved celestial datasets is proposed in this paper. In the first stage, a window contour detection method utilizing digital image processing techniques and comprising three operational steps: window extraction, window searching, and elliptical fitting, is employed to perform initial contour extraction. The second stage introduces an LPC-ResNet classifier, an adapted version of ResNet architecture, to supplement potential contour points missed by the window-contour detection method in the first stage. Experimental validation using New Horizons’ Long-Range Reconnaissance Imager (LORRI) images of Pluto and its biggest satellite Charon demonstrates that the window-based method achieves optimal Precision, indicating its superior capability in filtering non-contour pixels, yet exhibits limited Recall, likely due to unintended exclusion of contour points during elliptical fitting in the window-contour detection method. After incorporating the LPC-ResNet classifier, both Recall and <italic>F</italic>₁ score are improved, enhancing the ability to retain contour points and resulting in more complete contour extraction. To further validate the role of LPC-ResNet classifier in refining detection outcomes, centroid measurement experiment of Pluto and Charon is conducted using the contours extracted by the two-stage detection method and the single window-contour detection method, using the ephemerides Plu060 as the reference. The results of the experiment demonstrate that the two-stage method reduces mean and standard deviation of the Observed-minus-Calculated (O-C) residuals in both <italic>x</italic> and <italic>y</italic> directions, which means that the two-stage method has higher detection precision compared to the single window-contour detection method, confirming that LPC-ResNet classifier can improve the contour extraction effect of the window-contour detection method.