Personality Recognition Based on Palmistry Using DeepLearning YOLOv5 and YOLO-NAS

Abstract

As an intangible cultural heritage and traditional belief system, palmistry has been conventionally utilized as a medium to understand human personality. However, its interpretation remains subjective and manual, while previous digital studies have been restricted to single-object recognition. This study aims to develop an automated multi-object recognition system for palmistry features, which include palmar lines, mounts, fingers, and hand types, by employing YOLOv5 (anchor-based) and YOLO-NAS (anchor-free) architectures under the constraint of a small-scale dataset. The research phases encompass data selection integrated with data augmentation, bounding box annotation, data splitting, as well as model training and testing evaluated using Mean Average Precision (mAP), precision, and recall metrics. Experimental results demonstrate that YOLOv5 outperforms YOLO-NAS under limited data conditions, achieving a precision of 0.800 and an mAP of 0.846, compared to YOLO-NAS which yields a precision of 0.104 and an mAP of 0.603. Conversely, YOLO-NAS records a higher recall of 0.894. The imbalance between the recall and precision values in YOLO-NAS is primarily influenced by the limited training samples and the implementation of int8 quantization techniques. This study contributes by establishing the efficiency boundaries of deep learning architectures for the digitalization of cultural heritage based on limited datasets.