AGM-YOLO: A Generalizable Plug-and-Play Attention Module for Enhancing Small Object Detection Across Biomedical and Remote Sensing Imagery

The accurate detection of minute objects in biomedical images is a critical bottleneck for early disease diagnosis. These targets pose a significant challenge to existing algorithms because of their small size, low signal-to-noise ratio, and tendency to be obscured by complex backgrounds. Although general-purpose detection methods have advanced considerably, they still exhibit deficiencies in precision and recall for such diminutive targets. To address this, we designed and proposed a novel plug-and-play neural network module, termed AGM-YOLO, engineered to significantly enhance the detection capabilities of YOLO-family architectures. The core innovation of AGM-YOLO lies in its sophisticated synergistic multi-attention architecture, which combines a customized multi-head self-attention (CMHA), a lightweight channel attention block (GEBlock), and an efficient multiscale feature refinement neck (MFR-Neck). This design systematically increases the detection performance in multiple dimensions. Extensive experiments on three public datasets spanning the biomedical and remote sensing domains demonstrated that YOLO models equipped with AGM-YOLO substantially outperformed various state-of-the-art baselines. By setting a new benchmark for both precision and recall, AGM-YOLO shows considerable potential for advanced automated clinical diagnosis.