基於輕量化深度學習模型之愛文芒果不良品辨識研究

Abstract

愛文芒果是台灣具代表性的高經濟價值農產品之一，長年作為主要的外銷水 果，為台灣農業帶來可觀的出口收益與國際能見度。為了維持其品牌形象與市場 競爭力，採收後的品質檢查成為重要的一環。傳統上，芒果的品質判定多依賴人 工視覺檢查，不僅耗費人力與時間，亦可能因主觀判斷造成品質不一致的問題。 隨著人工智慧技術的快速發展，特別是電腦視覺與深度學習在圖像辨識領域的應 用日益成熟，將 AI 影像辨識導入農業生產線，已成為提升自動化與標準化品質檢 查流程的關鍵方向。過去已有研究使用深度學習模型，成功應用於愛文芒果不良 品分類任務上，達成相當程度的準確度，證明了 AI 在農業分級應用中的可行性。 然而，實務上我們面臨兩個主要挑戰:一是農業相關任務的標註影像資料取得不 易，導致可用資料量相對稀少;二是多數部署情境發生在具資源限制的邊緣設備上，對模型大小與推論速度有嚴格限制。因此，傳統大型模型雖表現優異，但實 際應用時往往因推論延遲、記憶體占用過高等問題而難以落地。 本研究即針對上述痛點，探討在圖像資料量有限的情況下，是否能透過輕量 化的 CNN 架構，兼顧模型效能與部署彈性。我們設計並實作數個參數量相對較 少的 CNN 模型，並與參數量較大的中大型模型進行全面比較，包括準確率、 macro-F1、參數量與推論速度等指標。此外，我們亦引入目前熱門的大型視覺語 言模型，例如 Gemma3 等作為對照組，探討其在資料量少、任務相對冷門的農業 情境下之表現與可行性。 實驗資料採用台灣 AI CUP 2020:愛文芒果不良品分類競賽所提供之公開資 料集，涵蓋多種瑕疵類型與不同品質等級的芒果影像。結果顯示，所提出之輕量 化模型在面對有限訓練資料時，仍可達到與大型模型相近甚至更穩定的分類表現， 且大幅降低模型參數量與推論資源需求。在與 VLM 比較上，我們亦觀察到即便 大型模型具備跨任務學習能力，但在專一任務、且訓練資料目標清晰的情況下， 輕量模型更具資源效率與應用潛力。 本研究驗證了小模型在單一領域場景下的實用性，並比較不同模型的效果， 不僅有助於農業智慧化系統的落地實施，也對深度學習模型的選擇策略提供新思 維。藉此推動採收後自動化處理的效率與品質一致性，提升台灣本地愛文芒果的 商品附加價值與品牌精緻化程度。

Irwin mango is one of Taiwan's most iconic and economically valuable agricultural products. As a major export fruit, it has long contributed significantly to Taiwan's agricultural export earnings and international visibility. To maintain its brand image and market competitiveness, post-harvest quality inspection plays a critical role. Traditionally, mango quality assessment relies heavily on manual visual inspection, which is labor- intensive and time-consuming, and may lead to inconsistent quality due to subjective judgment. With the rapid advancement of artificial intelligence, particularly in computer vision and deep learning for image recognition, incorporating AI-based image analysis into agricultural production lines has become a key direction for enhancing automation and standardization of quality inspection processes. Previous studies have successfully applied deep learning models to classify defective Irwin mangoes, achieving promising levels of accuracy, thereby demonstrating the feasibility of AI in agricultural grading applications. However, two major challenges remain in practice: (1) annotated image data for agricultural tasks is difficult to obtain, resulting in limited available training data; and (2) many deployment scenarios occur on resource-constrained edge devices, where strict limitations on model size and inference speed exist. As a result, although large-scale models perform well, their deployment is often hindered by latency and memory usage issues. This study addresses the above challenges by investigating whether lightweight CNN architectures can achieve a balance between model performance and deployment flexibility under limited image data conditions. We designed and implemented several CNN models with relatively fewer parameters and conducted comprehensive comparisons with larger models in terms of accuracy, macro-F1 score, parameter count, and inference speed. Additionally, we included state-of-the-art vision-language models (VLMs), such as Gemma3, as a baseline to explore their performance and feasibility in low-resource, domain-specific agricultural tasks. The experiments were conducted using the publicly available dataset from the Taiwan AI CUP 2020: Irwin Mango Defect Classification Challenge, which contains images of mangoes with various defect types and quality grades. Results show that our proposed lightweight models can achieve comparable or even more stable classification performance than larger models under limited training data, while significantly reducing parameter size and inference resource demands. When compared to VLMs, we also observed that although large models possess cross-task learning capabilities, lightweight models are more resource-efficient and practically applicable for specific tasks with clearly defined training objectives. This study validates the practicality of small models in niche domains and data- scarce scenarios. It not only contributes to the implementation of intelligent agricultural systems but also provides new insights into deep learning model selection strategies. By improving the efficiency and consistency of post-harvest processing, the proposed approach enhances the added value and branding quality of Taiwan's Irwin mangoes.