Abstract: In order to achieve rapid and accurate identification of key phenological periods of fruit trees using machine vision system in digital orchards, 15 000 images were collected from four phenological periods of four fruit trees such as apple, mango, pomegranate and citrus in Sichuan, and the training, validation and test data sets were randomly divided in a ratio of 6∶2∶2. Four deep learning classification models such as VGG16, ResNet50, MobileNetV2 and Swin Transformer were trained to evaluate the accuracy and performance of different models. The results showed that the accuracy of each model was 98.9%, 99.3%, 99.7%, 99.8%, respectively. The error of identifying phenology in maturation stage of mango was 96.7%, 98.2%, 99.0% and 99.5%, respectively. The computational complexity (GFLOPs) for model recognition on the test data were 15.50, 4.12, 0.32, and 15.14, respectively. The single image test was 3.00 ms, 2.33 ms and 3.00 ms, 4.67 ms, respectively. The results can provide reference for the selection model of machine vision system for orchard serverside and embedded edge devices.

Key words: fruit tree, phenological phase, deep learning, image classification, selfattention mechanism

摘要： 为实现数字果园的机器视觉系统快速准确识别果树关键物候期，采集四川地区苹果、杧果、石榴、柑橘4种果树4个物候期的图像15 000幅，按6∶2∶2的比例随机划分训练、验证和测试数据集，训练VGG16、ResNet50、MobileNetV2及Swin Transformer 4个深度学习图像分类模型，评测不同模型的精度和性能。结果表明，各模型识别物候期精度分别为98.9%、99.3%、99.7%、99.8%，其中杧果成熟期的识别误差较大，精度分别为96.7%、98.2%、99.0%、99.5%；模型识别测试集图像的计算量(GFLOPs)分别为15.50、4.12、0.32、15.14，识别单张图像耗时分别为3.00ms、2.33ms、3.00ms、4.67ms。该结果可为果园嵌入式设备、服务器端的机器视觉系统选择模型提供参考。

关键词: 果树, 物候期, 深度学习, 图像分类, 注意力机制