Small Object Detection Method for Bioimages Based on Improved YOLOv8n Model.

Abstract

As natural science research penetrates further into the microscopic world, the biological discipline has an increasing demand for tools to observe sub-micrometer structures such as cell structure and biomolecule assembly. Electron microscopy imaging has emerged as a pivotal method for such observations, yet accurate identification remains challenging due to the high density, mutual occlusion, small size, and diverse postures of the targets. To date, no research has systematically addressed these issues, limiting progress in biological microscopic research. Here, we introduce an improved YOLOv8n model for detecting the bird feather hooklet, a typical microscopic target within electron microscope images. The improved model incorporates three modules: gather-excite attention mechanism (global-local feature integration), explicit visual center (EVC) module (small-object detection enhancement through global and local feature fusion), and Shape IoU loss function (bounding-box regression optimization for posture variations). The experimental outcomes demonstrate that, compared to the baseline model, the improved YOLOv8n achieves a 3.5% increase in precision, a 9.1% boost in recall, and a 5.7% improvement in mAP50, along with 4.4% and 6.3% gains in mAP50-95 and F1 score, respectively. These advancements demonstrate the improved YOLOv8n model's effectiveness in detecting occluded, aggregated, and multi-posed hooklets at the nanometer level, offering new insights into feather structure-function relationships and advancing ornithological research. This study not only highlights the great potential of the improved YOLOv8n model in complex object detection but also emphasizes its application significance in micro-precision biological research.