Fine-Grained Palmar Morphology Detection: A Comparative Analysis of YOLOv5 and YOLOv8 Architectures for Palmistry Knowledge Representation

Abstract

Palmar morphology offers a unique and complex taxonomy of features; however, its analysis in traditional practices is still hindered by human subjectivity and inconsistent interpretation. While biometrics have advanced, the automated detection of 12 diverse and fine-grained palmar elements remains a significant challenge due to high spatial complexity. This study aims to establish a standardized digital framework for identifying 12 heterogeneous palmar morphologies, bridging the gap between traditional taxonomy and modern computer vision while supporting palmistry knowledge representation through structured and computationally reproducible morphological analysis. We propose a comparative benchmark between anchor-based (YOLOv5) and anchor-free (YOLOv8) architectures. Optimization strategies were implemented through transfer learning using MS COCO pre-trained weights to overcome small-scale dataset constraints and enhance multi-scale feature extraction on subtle biological markers. Experimental results demonstrate that YOLOv8-L is the superior model, achieving an mAP@0.5 of 0.9244, significantly outperforming YOLOv5 in localizing fine-grained features. The integration of C2f structures and the anchor-free paradigm proved effective in handling intricate spatial relationships, achieving a high inference speed of 97.2 FPS. This research presents an automated and standardized framework for biological feature extraction, enabling objective digital documentation and consistent interpretation of palmar morphological features. The findings also indicate that anchor-free models are more robust for detecting complex biological morphologies.