Green Apple Detection Method Based on Optimized YOLOv5 Under Orchard Environment

In orchards, detecting green apples can be challenging due to interference factors like similar fruit and background color, branch and leaf shading, and fruit overlap. To address this limitation, this paper presents a simple yet effective detection model based on improved YOLOv5, which can enhance the detection ability of green apples against a near-color background. Our contributions are twofold. Firstly, we added an attention mechanism to enhance the feature extraction network of the conventional YOLOv5. This modification focuses the network on green apple features and improves the detection performances on green apples. Secondly, we introduced the Focal Loss calculation method to the loss calculation of YOLOv5, to improve the model's results by controlling positive and negative sample weights as well as hard and easy to classify sample weights. Experimental results show that our model yields better results. While the base YOLOv5 model achieved an Average Precision (AP) of 86.3% and an Average Recall (AR) of 66.8% on the green apple dataset test, our improved YOLOv5 model reached an AP of 88.1% (a 1.8 percentage point improvement) and an AR of 69.1% (a 2.3 percentage point improvement). Our proposed model, therefore, significantly enhances detection efficiency.