COINS: Counting Cones Using Inpainting Based Self-supervised Learning

A novel approach for "COunting cones using IN-painting based Self-supervised learning (SSL)"(COINS), in wide field-of-view, low-resolution Adaptive Optics (AO) images is described. The proposed approach is applied to a dataset of 4°×4° AO images captured using an AO rtx1 device. The SSL model employed inpainting pretext task to learn representations, which were fine-tuned using a small number of expert annotated images. Our experimental results show that the proposed method significantly outperforms a baseline Delaunay triangulation Voronoi algorithm-based cone counting and can accurately count cones in 80×80 and larger regions anywhere within the field of view.