Multi-object sperm detection and tracking based on enhanced YOLOv4 and improved DeepSORT

Assisted reproductive technology plays a pivotal role in addressing infertility issues. Current computer-aided semen analysis systems face significant challenges in processing sperm microscopic images, particularly in detecting small sperm targets and handling indistinct sperm features. This study proposes an advanced multi-object sperm detection and tracking framework that integrates an enhanced YOLOv4 architecture with an optimized DeepSORT algorithm. Our method introduces a novel image slicing strategy to prevent small target loss and employs ResNet50 for improved feature extraction in tracking. Furthermore, a specialized sperm image augmentation strategy based on a heterogeneous Laplacian distribution model was implemented during model training. Experimental results demonstrate superior performance across various metrics for sperm detection and tracking. The system exhibits particularly robust performance in multiple small targets, outperforming existing state-of-the-art methods.