Artificial intelligence for celestial object census: the latest technology meets the oldest science

Large surveys using modern telescopes are producing images that are increasing exponentially in size and quality. Identifying objects in the generated images by visual recognition is time-consuming and labor-intensive, while classifying the extracted radio sources is even more challenging. To address these challenges, we develop a deep learning-based radio source detector, named \textsc{HeTu}, which is capable of rapidly identifying and classifying radio sources in an automated manner for both compact and extended radio sources. \textsc{HeTu} is based on a combination of a residual network (ResNet) and feature pyramid network (FPN). We classify radio sources into four classes based on their morphology. The training images are manually labeled and data augmentation methods are applied to solve the data imbalance between the different classes. \textsc{HeTu} automatically locates the radio sources in the images and assigns them to one of the four classes. The experiment on the testing dataset shows an average operation time of 5.4 millisecond per image and a precision of 99.4\% for compact point-like sources and 98.1\% for double-lobe sources. We applied \textsc{HeTu} to the images obtained from the GaLactic and the Galactic Extragalactic All-Object Murchison Wide-field Array (GLEAM) survey project. More than 96.9\% of the \textsc{HeTu}-detected compact sources are matched compared to the source finding software used in the GLEAM. We also detected and classified 2,298 extended sources (including Fanaroff-Riley type I and II sources, and core-jet sources) above $5\sigma$. The cross-matching rates of extended sources are higher than 97\%, showing excellent performance of \textsc{HeTu} in identifying extended radio sources. \textsc{HeTu} provides an efficient tool for radio source finding and classification and can be applied to other scientific fields.