A Comparative Machine Learning Study of YOLOv5–YOLOv11 for Ear Biometric Gender Classification

Resumen

Ear biometrics offers a practical, non-intrusive modality for gender classification and can support identification systems operating under challenging imaging conditions. In this work, multiple pretrained YOLO variants (YOLOv5–YOLOv11) were retrained via transfer learning for ear-based gender classification using a subset of the EarVN1.0 dataset consisting of 1,500 images (750 male and 750 female). Images were annotated and converted into the required YOLO training format, and all models were trained under consistent settings (learning rate 0.01 and batch size 16) on a GPU-enabled environment. Model performance was evaluated using precision, recall, and F1-score, with additional micro-, macro-, and weighted-F1 measures to assess overall effectiveness across classes; F1-score was used as a primary metric because it jointly reflects precision and recall. Experimental results show that YOLOv5 achieves the best overall performance among all tested models, obtaining Micro F1 = 0.83, Macro F1 = 0.79, and Weighted F1 = 0.88, while also maintaining a relatively compact retrained model size (42.23 MB). These findings indicate that YOLOv5 provides the most favorable performance–efficiency trade-off for ear-biometric gender classification within the studied setup.