基於 F2SA-YOLOv8在小黃瓜之偵測定位及採收

Abstract

在智慧農業領域，人工智慧視覺技術結合機械手臂已被視為解決農業人力短缺問題的有效方案。然而，作物遮蔽與辨識速度不足，仍然是自動化採收系統實用化的主要挑戰。針對上述問題，本研究提出一種融合模糊邏輯之注意力機制的F2SA-YOLOv8，應用於小黃瓜的辨識與定位，並設計兼具剪切與抓取功能的末端執行器，結合混合式視覺伺服控制策略，以提升整體採收成功率與定位精度。本研究所提出之F2SA-YOLOv8，導入自注意力（Self-Attention）機制並結合第二型模糊邏輯(Type-2 Fuzzy Logic)，顯著提升模型在複雜背景與作物遮蔽情境下的辨識能力。於自建農業資料集測試中，模型F1值達96.55%，平均推理時間縮短至256毫秒，分別較原始YOLOv8提升0.59 %與減少648毫秒，展現在雜環境中卓越的辨 識效率。在末端執行器設計方面，將攝影機安裝於剪切機構下方，並採用主動式視覺伺服策略，使執行器能於辨識與運動控制過程中主動推開遮蔽葉片，有效解決傳統採收機器人於作物遮蔽下的操作困難。控制策略部分，本研究融合基於影像的視覺伺服（IBVS）與基於位置的視覺伺服（PBVS），前者提供像素級的控制精度，後者則針對深度距離進行精確調整，即使在葉片遮蔽或角度變化下，仍能保持高度的穩定性與精度。與過往文獻相比，本研究於遮蔽場景下展現顯著改進，證實所提出之整合系統更適用於實際農業應用場景。整體成果可應用於溫室小黃瓜等條狀瓜果類作物的自動採收，不僅提升辨識與採收效率，亦具備推廣至其他作物的潛力。In smart agriculture, combining AI vision with robotic arms is a promising solution to labor shortages. However, crop occlusion and limited recognition speed remain key challenges for automated harvesting systems. This study proposes a Type-2 Fuzzy Self-Attention YOLOv8 (F2SA-YOLOv8) model for cucumber detection and localization, along with a dual-function end-effector that integrates cutting and gripping. A hybrid visual servo control strategy is adopted to improve harvesting success and positioning accuracy.The F2SA-YOLOv8 model integrates a self-attention mechanism with Type-2 fuzzy logic, significantly enhancing detection in complex and occluded environments. On a custom agricultural dataset, it achieved a 96.55% F1-score and an average inference time of 256 ms, 0.59% higher accuracy, and 648 ms faster than the YOLOv8, demonstrating excellent efficiency in cluttered scenes. The camera is mounted beneath the cutting mechanism, enabling active visual servoing to push aside leaves during detection and movement, effectively resolving occlusion issues. The control strategy combines Image-Based Visual Servoing (IBVS) for pixel-level alignment and Position-Based Visual Servoing (PBVS) for depth correction, maintaining high accuracy even under occlusion or varying stem angles.Compared with previous methods, the proposed system shows significant improvement in occluded conditions and is well-suited for greenhouse cucumber harvesting. It enhances both recognition and operational efficiency, and presents strong potential for extension to other crops.