針對小資料集之資料分類器增強輔助設計:以海龜為例

Abstract

本研究針對臺灣海域海龜個體辨識問題，提出一套資料增強輔助之三層級分類系統，以逐層分類方式協助縮小比對範圍、加速辨識流程。研究採用 YOLOv5、YOLOv7 與 YOLOv8 三種深度學習架構，分別訓練於物種分類（ClassA）、左眼下鱗片數量分類（ClassB）及臉部特徵分類（ClassC）等任務，並進行模型效能比較。實驗結果顯示，YOLOv8 於複雜特徵辨識任務（ClassC）中表現最佳，具備較佳的泛化能力與辨識穩定性。本研究所提出之系統具備良好之可擴充性及實務應用潛力，為未來自動化辨識系統建置提供實證基礎。

This study addresses the individual identification of sea turtles in Taiwan by proposing a three-tier classification system with data augmentation to assist in narrowing down candidate lists and accelerating the identification processes. The research employs three YOLO-based deep learning architectures—YOLOv5, YOLOv7, and YOLOv8—trained separately on species classification (ClassA), scute number under left eye classification (ClassB), and left facial scute pattern (ClassC) tasks, respectively. Model performance comparisons reveal that YOLOv8 achieves superior results in complex feature recognition tasks (ClassC), demonstrating better generalization and classification stability. Overall, the proposed system exhibits excellent scalability and practical potential, providing an empirical foundation for future automated identification systems.