A Comparative Study of YOLO Models for Sperm and Impurity Detection Based on Proposed Augmentation in Small Dataset

Sperm detection can help with a lot of different kinds of research, such as forensics, assisted reproduction, and fertility assessment. However, it can be challenging to detect the sperm, particularly when the sample size of the dataset is minimal. This paper presents a comparison of different YOLO (You Only Look Once) models for sperm and impurity detection based on the proposed augmentation in the small dataset. According to the results, the YOLOv5 model had high precision, recall, F1-score, and mean average precision (mAP) for training with both the original dataset and the proposed augmentation dataset. While the YOLOv8s model had the highest precision, recall, F1-score, and
$m$
AP (0.994, 0.965, 0.979, and 0.993, respectively), only training with the proposed augmentation. In short, when trained with the proposed augmentation, the YOLO models can detect sperm and impurity even when both the objects and the number of samples are small.