近年來，隨著環保意識的提升與資源回收作業量的增加，傳統的人工酒瓶分類與瑕疵檢測作業逐漸面臨效率與準確度的挑戰，因此，發展自動化分類技術是提升處理效率與降低成本的關鍵。目前，酒瓶等回收物的分類（例如品牌或類型辨識）與品質檢測（例如是否破損）仍高度依賴人工篩選，此過程不僅耗時、成本高昂，且難以維持一致的檢測標準。研究旨在開發一套自動化檢測系統以取代人工分類。此系統面臨的挑戰是需同時處理高層次的「語意分類」與底層的「瑕疵檢測」。為此，本研究的核心在於評估不同視覺語言模型（Vision-Language Model， VLM）策略的效能，並與高效的卷積神經網路（Convolutional Neural Network， CNN）在兩項任務中的適用性與效能差異。在分類任務上，我們比較了原始 CLIP、AnomalyCLIP 及採用CoOp+ATPrompt提示學習 (Prompt Learning) 策略的模型。在瑕疵檢測任務上，則重點比較了 AnomalyCLIP、CoOp+ATPrompt 與 ResNet50 模型的表現。實驗結果明確顯示，經過提示學習優化的 VLM 展現了全面的優勢。 在「分類任務」中，CoOp+ATPrompt 的表現顯著優於原始 CLIP 與 AnomalyCLIP。更重要的是，在「瑕疵檢測任務」中，CoOp+ATPrompt 同樣展現了最佳效能，其辨識率超越了傳統的 ResNet50 模型與 AnomalyCLIP。基於上述發現，本研究證實採用 CoOp+ATPrompt 的視覺語言模型可作為一個統一的解決方案，同時高效執行酒瓶分類與瑕疵檢測任務。此結果顯示，先進的提示學習策略能顯著提升 VLM 在細粒度 (fine-grained) 檢測任務上的能力，使其表現超越了在此類任務中被視為強項的 CNN 模型。 本研究最終建構了一個基於 CoOp+ATPrompt 的自動化辨識系統，為未來智慧回收與自動化檢測系統的設計提供了一個更簡潔且高效能的架構參考。 關鍵詞：自動化回收、酒瓶分類、瑕疵檢測、深度學習、提示學習、視覺語言模型 (VLM)、卷積神經網路 (CNN);In recent years， with the rise of environmental awareness and the increasing volume of resource recycling， traditional manual bottle classification and defect detection have faced significant challenges in efficiency and accuracy. Developing automated classification technologies has become key to enhancing processing efficiency and reducing costs. Currently， the classification (e.g.， brand or type identification) and quality inspection (e.g.， breakage detection) of recyclables like glass bottles still rely heavily on manual sorting. This process is not only time-consuming and costly but also difficult to maintain consistent inspection standards.This study aims to develop an automated inspection system to replace manual sorting. The challenge lies in simultaneously handling high-level "semantic classification" and low-level "defect detection." To address this， the core of this research is to evaluate the performance of different Vision-Language Model (VLM) strategies and compare their applicability and efficacy against efficient Convolutional Neural Networks (CNN) in both tasks. For the classification task， we compared the original CLIP， AnomalyCLIP， and models employing the CoOp+ATPrompt prompt learning strategy. For the defect detection task， the performance of AnomalyCLIP， CoOp+ATPrompt， and ResNet50 was evaluated.Experimental results clearly demonstrate that the VLM optimized with prompt learning exhibits comprehensive advantages. In the "Classification Task，" CoOp+ATPrompt significantly outperformed the original CLIP and AnomalyCLIP. More importantly， in the "Defect Detection Task，" CoOp+ATPrompt also achieved the best performance， surpassing the traditional ResNet50 model and AnomalyCLIP.Based on these findings， this study confirms that a VLM employing CoOp+ATPrompt can serve as a unified solution to efficiently execute both bottle classification and defect detection tasks. The results indicate that advanced prompt learning strategies can significantly enhance VLM capabilities in fine-grained detection tasks， enabling them to outperform CNN models typically considered dominant in such areas. Finally， this study constructs an automated recognition system based on CoOp+ATPrompt， providing a concise and high-performance architectural reference for the future design of smart recycling and automated inspection systems. Keywords: Automated Recycling， Bottle Classification， Defect Detection， Deep Learning， Prompt Learning， Vision-Language Model (VLM)， Convolutional Neural Network (CNN)
