Unity doubles the power: Modeling the optical spectrum (SDSS) via the generated neural network (GaSNet-III) and Searching LSB galaxies in KiDS with Deep Learning

We developed a novel generated neural network to reconstruct the rest-frame spectra by giving the observed spectra and their flux error. This network provided all the necessary information for modeling spectra, including the eigenspectra and coefficient. Using this reconstruction, we can achieve the classifying, redshift estimation, and anomaly detection in the same framework. Our test demonstrates we reach the same level of accuracy in spectral fitting and redshift estimation as the classical method, with at least O(10^(-2)) faster. Combining our previous work (GaSNet-II), which used sub-network assembly, we can achieve error estimation and subclassify for future spectroscopy survey (4MOST) pipelines.

and

Detecting and characterizing low-surface brightness galaxies (LSBGs) and ultra-diffuse galaxies (UDGs) is known to be challenging due to their faint surface brightness, posing a significant hurdle for traditional detection methods. Recently, artificial neural networks proven to have a powerful learning ability, which can learn the features of the object from the image and complete classification or parameter regression tasks. Thus, artificial neural networks have been increasingly applied in astronomy field to handle the rising volume and complexity of photometric images. Object detection algorithms, one of the branches of computer vision, have greater capabilities than traditional classification and regression algorithms. Thus, it can accomplish large-scale galaxy detection tasks. In this work, we propose LSBGnet framework, a deep neural network specifically designed for automatic detection of LSBGs and UDGs. First, we use the images from Sloan Digital Sky Survey (SDSS) to train and test the model. The performance of the LSBGnet-SDSS model is outstanding in this work, and the recall and precision of LSBGnet model is more than 97% on the test set. Then, given the excellent performance of LSBGnet, we decided to use it for large-scale detection for UDGs, a subset of LSBG, on KiDS DR5 (Kilo Degree Survey Data Release 5). We built the LSBGnet-KiDS model using the LSBGnet framework with an iterative detection method. We utilized the LSBGnet-KiDS model to detect UDGs from all photometric images of KiDS DR5 and obtained 966 UDG candidates. In this process, we successfully completed a large-scale detection for UDGs without using known UDG samples. When faced with a large-scale specific object detection task and the number of samples is not enough to build a model, we can utilize this method to handle it. It also provides an effective approach to detection for specific objects for the upcoming surveys.