近年來，由於其廣泛適用性，越來越多研究者專注於半監督學習，特別是多標籤分類。然而，多標籤影像分類並不容易，因為偽標籤的雜訊與語意重疊使學習過程難以充分發揮未標註資料的潛力。本論文提出一套新穎的類別自適應基於 Medoid 的對比學習（CAMCL）框架，結合分布校正、語意過濾與對比學習，以增強複雜多標籤關係下的語意意識。CAMCL包含三個階段。首先，透過不確定性一致性學習建立初始表徵。其次，引入了原型 Medoid 記憶庫用於偽標籤過濾，並結合類別自適應閾值與對比策略以精煉特徵空間。最後，利用保留的標記資料微調模型，進一步提升分類效能。實驗結果顯示，CAMCL 在分類效能與偽標籤可靠性方面均優於現有方法，為標註有限的情境下提供可擴展的解決方案。;Recently， owing to the widespread applicability， more and more researchers focus on semi-supervised learning techniques， especially for multi-label classification. However， multi-label image classification is not an easy task， since the pseudo-label noise and semantic overlap making the learning process difficult to fully exploit the potential of unlabeled data. In this study， a novel Class-Adaptive Medoid-Based Contrastive Learning (CAMCL) framework is proposed by integrating distribution calibration， semantic filtering， and contrastive learning to enhance semantic awareness under complex multi-label relationships. CAMCL consists of three stages. First， an initial representation is established through uncertainty-aware consistency learning. Second， a Prototype-Medoid memory bank is introduced for pseudo-label filtering， together with class-adaptive thresholds and contrastive strategies to refine the feature space. Finally， the model is fine-tuned using the reserved labeled data to further improve classification performance. Experimental results show that CAMCL outperforms existing methods in both classification performance and pseudo-label reliability， providing a scalable solution for annotation-limited scenarios.
