Abstract
Accurate and quantitative detection of boar sperm heads is essential for breeding selection and reproductive management. Manual microscopic counting is time-consuming, labor-intensive, and prone to subjective bias, while existing computer-based algorithms often struggle to recognize sperm cells accurately when they overlap or move rapidly in high-magnification microscopic images. This study proposes a lightweight boar sperm detection model, YOLO11_SRP, designed to improve small-object recognition in complex microscopic scenarios. The model integrates a lightweight StarNet backbone, a rectangular self-calibration module for enhanced spatial feature modeling, and an additional low-level detection layer optimized for tiny targets. We evaluated the model on a boar sperm microscopic image dataset and compared it with the standard YOLO11s framework. The results show that YOLO11_SRP achieves an mAP@0.5 of 91.9%, representing a 13.9% improvement over YOLO11s, while simultaneously reducing parameters by 39% and computational cost by 14.1%. These findings demonstrate that YOLO11_SRP provides efficient and accurate sperm detection, supporting the development of efficient and reliable automated sperm analysis pipelines, in which sperm head detection serves as a fundamental preprocessing step.