茄科作物成熟度分類模型的深度學習與色彩空間轉換研究

Abstract

番茄是世界上重要的經濟作物之一，其成熟度判斷對於收穫時機的選擇和市場價值的提升具有關鍵作用，且其品質會直接影響銷售的數量以及其營養成分的重要因素。以一位消費者的角度來說，其成熟程度是衡量其品質的重要指標，因此，自動化的成熟度評估是一個至關重要的研究主題，因為其對確保優質產品的最佳產量有一定程度上的幫助，進而提高收入。本研究旨在運用深度學習與色彩空間轉換的技術來建構針對茄科作物成熟度判斷模型，並以番茄為實驗對象。我們採用 YOLOv11 偵測模型定位番茄，接著使用經 PyTorch 訓練以 ResNet-50 為基礎的模型並結合了色彩空間轉換上的應用並辨識其成熟程度。資料集涵蓋不同生長階段的番茄影像，影像涵蓋了未成熟、半成熟、成熟不同階段。並透過旋轉、縮放、對比變化以及亮度變化等擴增方式提升模型泛化能力。為提取更具判別力的色彩特徵，我們進行了 RGB 至 HSV與 CIELAB 色彩空間轉換。實驗結果顯示，模型表現優異且在成熟度分類上的準確率可達九成，色彩空間與深度學習整合可有效應用於智慧農業領域。

Tomatoes are a major global economic crop. Assessing their ripeness is crucial for selecting the optimal harvest time and increasing market value. Ripeness directly impacts quality, which in turn affects sales volume and nutritional content. From a consumer’s perspective, ripeness is a key indicator of quality. Therefore, automated ripeness assessment is a critical research topic, as it helps to ensure the best possible yield of high-quality products and, consequently, boosts revenue. This study aims to build a ripeness assessment model for solanaceous crops, using tomatoes as the subject, by leveraging deep learning and color space conversion techniques. We used the YOLOv11 detection model to locate tomatoes. Following this, we employed a PyTorch-trained ResNet-50-based model, which integrated color space conversion, to identify their ripeness level. The dataset includes images of tomatoes at different growth stages: unripe, semi-ripe, and ripe. We used data augmentation techniques like rotation, scaling, and adjustments to contrast and brightness to enhance the model's generalization ability. To extract more discriminative color features, we converted images from the RGB color space to both HSV and CIELAB. The experimental results show that the model performs exceptionally well, achieving a classification accuracy of over 90% for ripeness levels. This demonstrates that integrating color spaces with deep learning can be effectively applied in the field of smart agriculture.