Maturity detection of ‘Huping’ jujube fruits in natural environment using YOLO-FHLD

To intelligently detect the maturity of ‘Huping’ jujube in the natural environment, this study proposed a lightweight ‘Huping’ jujube maturity detection method based on the YOLO-FHLD. Firstly, the C2fF module was introduced to the YOLOv8n model to create a lightweight model and enhance the extraction ability of ‘Huping’ jujube features. Secondly, a new feature fusion module HS-FPAN was used to improve the expression and detection accuracy of ‘Huping’ jujube with different levels of maturity. Then, Focal Loss was employed as a loss function to address the class imbalance problem. Finally, knowledge distillation strategy further enhanced the detection accuracy of the model. The experimental results showed that the F1 score and mean average precision (mAP) of YOLO-FHLD increased by 1.24 % (reaching 79.48 %) and 2.96 % (reaching 85.40 %) compared with YOLOv8n, respectively. Furthermore, the model size was reduced to a mere 3.51MB, which accounted for only 58.99 % of the original model's size. In terms of Classification Error (Cls), Localization Error (Loc), and Background Error (Bkg), YOLO-FHLD decreased by 0.9, 0.68, and 0.38, respectively. For the ‘Huping’ jujubes during harvestable maturity period (cs) and non-harvestable maturity period (ws), the average precision (AP) of ‘Huping’ jujube detection reached 89.20 % (increased by 2.85 %) and 81.60 % (increased by 3.07 %), respectively. Compared to Faster R-CNN, SSD, YOLOv3-Tiny, YOLOv4-Tiny, YOLOv5n, YOLOv7-Tiny, and YOLOv8s models, the mAP of YOLO-FHLD increased by 24.84 %, 4.72 %, 7.35 %, 1.3 %, 2.38 %, 8.2 %, and 0.79 %, respectively. The model size was only 3.24 %, 7.67 %, 21.09 %, 16.56 %, 94.61 %, 29.97 % and 16.36 % of the above seven models , respectively. Consequently, this model has advantages in both detection accuracy and model size. This study can provide a methodological support for target detection of ‘Huping’ jujube fruits in natural environment.