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

Abstract

The detection and search of Low Surface Brightness Galaxies (LSBGs) is an object detection task based on astronomical observational imagery, aiming to accurately localize faint and diffuse galaxies from massive sky survey datasets. Due to the inherently low signal-to-noise ratio (SNR) and weak galaxy signals in the images, LSBG detection poses significant challenges. In this case, existing object detection methods struggle to effectively address issues such as feature degradation, semantic isolation, and background interference, often resulting in missed or false detections. To address these challenges, we propose a detection network termed DAFF-Net, based on Dynamic Adaptive Feature Fusion. The network is built upon a multi-scale Triangular Dynamic Neck (Tri-Neck), which integrates lateral and downsampling connections to construct a dual-path fusion architecture. This design enables efficient cross-level feature integration, thereby mitigating the issues of degraded features and semantic discontinuity in small-scale LSBG objects. To further enhance object representation capability, we embed a Dynamic Channel Attention (DCA) module into each pathway of the Tri-Neck. By jointly leveraging channel and spatial information, the DCA module adaptively focuses on critical regions of LSBG objects. Besides, we introduce an Implicit Intersection-over-Union (IIOU) loss function that incorporates corner distance and shape constraints, significantly improving the geometric alignment and regression robustness of bounding boxes. Experimental results demonstrate that DAFF-Net achieves an Average Precision (AP) of 95.62% and an AP for small objects (APs) of 31.37% on the SDSS dataset, outperforming mainstream detection models such as Faster R-CNN and the YOLO series. Moreover, DAFF-Net successfully identifies 765 candidate LSBGs in real SDSS observations. The generalizability of the Tri-Neck structure is further validated on the COCO 2017 benchmark dataset, achieving an AP of 40.0%.