Detection and classification of astronomical sources with Astro-RetinaNet in crowded stellar fields

TL;DR

A robust deep learning model, Astro-RetinaNet, based on the Retinanet algorithm to detect and classify blended sources in single-band astronomical images, significantly outperforming Photutils and SExtractor by factors of 3.4 and 7.1, respectively.

Abstract

Upcoming next-generation sky surveys will detect large number of faint objects with magnitudes larger than 25. When objects are crowded within a limited a field of view, blending becomes unavoidable. Blending leads to the omission of many sources during photometry in these fields, which cause an underestimates of tens of percent in crowded fields, and remains a major challenge for existing source-extraction techniques. Although artificial neural networks had shown promising results in the detection and classification in wide-field surveys, they often fail with severely blended stars. We developed a robust deep learning model, Astro-RetinaNet, based on the Retinanet algorithm to detect and classify blended sources in single-band astronomical images. After training and evaluating the performance of our network on simulated images, we find precision of 0.96, 0.89,0.70, 0.50,0.75 for single star, 2-star, 3-star, 4-star and 5-or-more star blending cases, respectively, with star number density $\sim$22000 stars per $\rm arcmin^2$. We compare our method's detection capability and completeness both on CSST simulated NGC 2298 images and HST observed M31 images. In crowded and non-crowded stellar fields of simulated NGC 2298, our results show that the model can recover $82\%$ and $95\%$ sources respectively at magnitude ($i$ band) of 25, while for SExtractor and Photutils the completeness reduces to $20\%, 59\%$ and $60\%, 88\%$ respectively. In the M31 case, as faint as 27 magnitude ($F814W$) in a crowded field, Astro-RetinaNet detects 2,224 sources, significantly outperforming Photutils and SExtractor by factors of 3.4 and 7.1, respectively.

Figures (11)

• Figure 1: Typical example of GalSim CSST training set. Pixel coordinates and magnitudes were sampled from Eq. \ref{['eq1']} and Eq. \ref{['eq:2']}. The pixel coordinates for each star are denoted by blue circles, and their corresponding bounding boxes are represented by red squares, both directly labeled. The numerical value appearing in the top-right corner of each box indicates the blending type (labels).
• Figure 3: Left: Magnitude distribution of mock catalog, with the blue histogram showing observed power-law counts and the yellow curve indicating the power-law probability density function (PDF). Right: A typical example of the distribution of blending categories in a single training image, where single stars (twice as frequent as blended categories) dominate to prevent misclassification, while other blending categories follow a uniform distribution.
• Figure 4: Reprinted from Tsung-YiLin_2017, with permission from Tsung-Yi Lin. The one-stage RetinaNet network architecture uses a Feature Pyramid Network (FPN) backbone on top of a feedforward ResNet architecture (a) to generate a rich, multi-scale convolutional feature pyramid (b). To this backbone RetinaNet attaches two subnetworks, one for classifying anchor boxes (c) and one for regressing from anchor boxes to ground truth object boxes (d). The network design is intentionally simple, which enables this work to focus on a novel focal loss function that eliminates the accuracy gap between our one-stage detector and state-of-the-art two-stage detectors like Faster R-CNN with FPN while running at faster speeds..
• Figure 5: Loss function curves during training (solid blue line with red circles) and validation (solid origin line with blue circles). The model converges at 50 epochs with a total training duration of 100 epochs.
• Figure 6: Diagram of confusion matrix of classification capability tested in test data set, over 140,000 sources(at least 85% of sources are blending stars) involved in this statistics. The darker the color in the diagonal, the better the performance of the classifier.