基于YOLO-L的自然環境中澳洲堅果果實的檢測和識別

摘要： 針對自然環境下果實重疊、相互遮擋和目標小的澳洲堅果果實檢測準確率低的問題，提出一種改進YOLOv9模型的識別方法（YOLO-L）。首先，引入BiFormer注意力機制，該機制通過雙層路由注意力機制實現了動態、查詢感知的稀疏注意力分配，能夠很好地捕獲特征表征，增強網絡對全局特征的關注度；其次，采用VoVGSCSP模塊代替YOLOv9中的CBFuse模塊，提高了復雜場景下小目標的檢測效果；最后，將YOLOv9模型默認的損失函數替換成排斥損失函數，解決了果實排列密集和漏檢的問題，進一步提升了澳洲堅果果實檢測的平均精度。通過消融試驗和對比試驗來驗證模型的有效性，發現YOLO-L模型的平均精度均值、精確率、召回率和F₁值分別達到96.2%、92.3%、88.2%和90.2%。與YOLOv9模型相比，YOLO-L模型的平均精度均值提升了4.9個百分點。總體而言，YOLO-L模型能夠在自然環境下準確識別被遮擋、重疊的澳洲堅果果實，且檢測精度高。研究結果可為澳洲堅果產業的智能采摘提供有效的技術支持。

關鍵詞： 圖像處理；深度學習；YOLOv9模型；澳洲堅果

中圖分類號： TP301.6；S664 文獻標識碼： A 文章編號： 1000-4440（2024）11-2102-09

Macadamia （Macadamia integrifolia Maiden amp; Betche） detection and recognition in natural environments based on YOLO-L

LIN Zuxiang¹， WANG Yingdong¹， MA Rong¹， WEI Yunsong¹， LI Ziwen¹， LI Jiaqiang¹， HE Chao²

（1.College of Mechanical and Transportation Engineering， Southwest Forestry University， Kunming 650224， China;2.Dehong Vocational College， Dehong 678400， China）

Abstract： Aiming at the issue of low detection accuracy for macadamia nuts in natural environments due to overlapping， mutual occlusion， and small targets， an improved YOLOv9 model recognition method （YOLO-L） was proposed. Firstly， the BiFormer attention mechanism was introduced， which achieved dynamic and query-aware sparse attention allocation through the Bi-level routing attention mechanism. This mechanism was capable of effectively capturing feature representations and enhanced the network’s focus on global features. Secondly， the VoVGSCSP module was used to replace the CBFuse module in YOLOv9， which improved the detection performance for small targets in complex scenes. Lastly， the default loss function of the YOLOv9 model was replaced with an exclusion loss function， which solved the problems of dense fruit arrangement and missed detections， and further enhanced the average accuracy of macadamia nut detection. The effectiveness of the model was validated through ablation and comparative experiments. It was found that the mean average precision， precision， recall， and F₁ score of YOLO-L model reached 96.2%， 92.3%， 88.2%， and 90.2%， respectively. Compared with the YOLOv9 model， the mean average precision of the YOLO-L model was improved by 4.9 percentage points. Overall， the YOLO-L model can accurately identify occluded and overlapped macadamia nuts in natural environments with high detection accuracy. The research results can provide effective technical support for the intelligent harvesting in the macadamia industry.

Key words： image processing;deep learning;YOLOv9 model;macadamia （Macadamia integrifolia Maiden amp; Betche）

澳洲堅果（Macadamia integrifolia）被譽為“堅果之王”，其果仁含有17種氨基酸、多種礦物質和大量蛋白質，是經濟價值較高的食用干果之一^[1]。然而，澳洲堅果采摘過程依賴人工，效率低、速度慢且勞動強度大，因此自動化和智能設備化采摘是解決該問題的重要途徑，但是復雜的生長環境和密集重疊的果實會影響視覺檢測的準確性，造成智能采摘進展緩慢。因此，快速準確識別澳洲堅果果實是智能采摘的關鍵。……