EHS-YOLO Model for Saccharomyces Cerevisiae Cell Detection

This study proposes a novel Enhanced High-level Screening-feature-YOLO (EHS-YOLO) model to address challenges in the Saccharomyces cerevisiae cell detection, including morphological similarities between budding cells, background impurities, and dense cell distribution. First, the EHS-YOLO model designs Enhanced Multi-scale Pooling Pyramid Module (EMSPPM) structure, which reduces the interference of broken bubbles and background impurities during model training. Second, the neck structure is replaced with an Efficient High-level Screening-feature Feature Pyramid Network (EHS-FPN) structure, enabling the model to efficiently integrate low-level detail information with high-level semantic information. Additionally, an attention mechanism is employed to enhance information interaction between different channels, allowing the model to focus more on cell structures and boundaries. Finally, the introduction of Powerful Intersection over Union (PIoU) addresses the issue of anchor box enlargement, prompting the model to pay closer attention to cell sizes and thereby improving algorithm performance. Compared to the YOLOv8n algorithm, EHS-YOLO has improved the detection accuracy [email protected] by 4.8%, recall by 6.6%, and mAP50 by 3.3%, while also reducing the number of parameters by 28.5% and the number of floating-point operations by 12.4%.