船舶熱成像分類之深度學習模型與資料集分割策略

Abstract

本研究針對船舶熱成像影像分類任務，探討不同深度學習模型與資料分割策略對分類效能之影響。由於熱成像影像能在夜間或低能見度環境下提供穩定的觀測能力，對於海上監測與智慧航行具有重要應用價值。然而，現有的相關研究與資料集仍相對不足。 為此，本研究建置一套涵蓋貨船、渡輪、遊艇及其他船舶四個類別之熱成像資料集，透過資料處理、分類標註與分層抽樣，共獲得4,834張有效影像，並採用兩種資料集分割策略進行模型訓練與測試。實驗選用基於卷積神經網路架構的YOLOv8-cls模型，以及基於Transformer架構的ViT模型進行比較。 結果顯示，在本研究所使用之資料規模與特性下，YOLOv8-cls在收斂速度與分類穩定性方面呈現較佳表現；而ViT模型對資料量與多樣性較為敏感，當資料有限時其優勢較不明顯，需在資料量或影像變化更充足的情況下，才有機會充分展現其全域特徵建模能力。此結果顯示不同模型在不同資料情境下的適用性差異。 本研究提供完整的資料集建置流程與模型效能比較，建立可重複、可擴充的實證資源，為後續船舶熱成像影像分類及相關應用研究提供參考。未來可進一步朝向多模態資料融合、混合式架構設計與智慧航行應用等方向發展。

This study investigates the influence of different deep learning model architectures and dataset partitioning strategies on ship thermal image classification performance. Thermal imaging provides stable observation capabilities under nighttime or low-visibility conditions, making it invaluable for maritime surveillance and intelligent navigation. However, existing research and publicly available datasets remain limited, restricting the development of deep learning applications in this domain. To address this gap, this study constructs a dedicated thermal image dataset comprising four vessel categories: cargo ship, ferry, yacht, and other ship. Through rigorous data cleaning, classification annotation, and stratified sampling, a total of 4,834 valid images were obtained and used for model training and evaluation under two dataset partitioning strategies. Two representative architectures were selected for comparison: the convolutional neural network-based YOLOv8-cls model and the self-attention-based Vision Transformer (ViT) model. Experimental results show that, under the data scale and characteristics used in this study, YOLOv8-cls demonstrates faster convergence and more stable classification performance. In contrast, the ViT model is more sensitive to dataset size and diversity; when data are limited, its advantages are less pronounced and may only emerge when larger or more varied datasets are available, enabling its global feature modeling capabilities. These findings suggest that the applicability of different models varies depending on data conditions, rather than indicating absolute performance superiority. This research provides a comprehensive dataset construction process and model performance comparison, establishing a reproducible and extensible empirical resource for future studies. The findings can serve as a reference for subsequent research in ship thermal image classification and related applications. Future work may further explore multimodal data fusion, hybrid model architectures, and applications in intelligent maritime navigation.