DAFF-Net: A detection and search method for small-scale low surface brightness galaxies.

Area of Science:

• Astronomy and Astrophysics
• Computer Vision
• Machine Learning

Background:

• Low Surface Brightness Galaxies (LSBGs) are difficult to detect due to low signal-to-noise ratios.
• Existing object detection methods struggle with feature degradation and background interference for LSBG detection.

Purpose of the Study:

• To propose an effective deep learning network, DAFF-Net, for accurate LSBG detection.
• To address challenges in LSBG identification, including faint signals and complex backgrounds.

Main Methods:

• Developed DAFF-Net featuring a multi-scale Triangular Dynamic Neck (Tri-Neck) for efficient feature fusion.
• Integrated a Dynamic Channel Attention (DCA) module to enhance object representation.
• Introduced an Implicit Intersection-over-Union (IIOU) loss function for improved bounding box regression.

Main Results:

• DAFF-Net achieved 95.62% Average Precision (AP) and 31.37% AP for small objects (APs) on the SDSS dataset.
• Successfully identified 765 candidate LSBGs in SDSS observations.
• The Tri-Neck structure achieved 40.0% AP on the COCO 2017 benchmark, demonstrating generalizability.

Conclusions:

• DAFF-Net significantly outperforms existing models in LSBG detection.
• The proposed Tri-Neck architecture and attention mechanisms enhance the detection of faint astronomical objects.
• DAFF-Net offers a robust solution for discovering LSBGs in large astronomical surveys.