Authors

Abstract

The intelligent detection of young apple fruits based on deep learning faced various challenges such as varying scale sizes and colors similar to the background, which increased the risk of misdetection or missed detection. To effectively address these issues, a method for young apple fruit detection based on improved YOLOv5 was proposed in this paper. Firstly, a young apple fruits dataset was established. Subsequently, a prediction layer was added to the detection head of the model, and four layers of CA attention mechanism were integrated into the detection neck (Neck). Additionally, the GIOU function was introduced as the model's loss function to enhance its overall detection performance. The accuracy on the validation dataset reached 94.6%, with an average precision of 82.2%. Compared with YOLOv3, YOLOv4, and the original YOLOv5 detection methods, the accuracy increased by 0.4%, 1.3%, and 4.6% respectively, while the average precision increased by 0.9%, 1.6%, and 1.2% respectively. The experiments demonstrated that the algorithm effectively recognized young apple fruits in complex scenes while meeting real-time detection requirements, providing support for intelligent apple orchard management.

Abstract in Chinese

基于深度学习的苹果幼果智能化检测面临诸多挑战如尺度大小不一、颜色与背景相近等，会导致误检或漏检的风险增加。为了有效解决这些问题，本文提出一种基于改进YOLOv5 的苹果幼果检测方法，首先建立苹果幼果数据集，再者在检测模型的检测头中添加预测层，在检测脖颈（Neck）中添加四层 CA 注意力机制，并引入GIOU 函数作为模型的损失函数，以提高模型的整体检测性能。在验证数据集上的准确率达到 94.6%，平均精度为 82.2%；与 YOLOv3、YOLOv4 和原始的 YOLOv5 检测方法相比，准确率分别提升 0.4%、1.3%、4.6%，平均精度分别提升 0.9%、1.6%、1.2%。试验证明，该算法能在满足实时检测要求的前提下，能够有效地识别复杂场景中的苹果幼果，为智能化苹果园管理提供支持。