基于多器官特征融合的枣品种识别方法

针对自然场景下的枣品种识别问题，以枣果为研究对象的机器视觉技术已成为枣品种精准识别的主流方法之一。针对枣品种存在类间差异小、类内差异大的问题，提出了一种基于多器官特征融合的枣品种识别方法。首先利用YOLO v3检测算法将采集的自然场景图像中的枣果和叶片器官分割提取，提出了基于笛卡尔乘积构建两器官组合对的枣品种多样本数据集，然后基于EfficientNetV2网络模型，设计了能够充分学习两器官特征相关性的融合策略来提升模型性能，引入了逐步迁移训练方式以提升枣品种识别效率。最后，在构建的包含20个枣品种数据集上进行了大量实验，得到97.04%的识别准确率，明显优于现有研究结果，并且在训练时间和收敛速度上，本方法也有一定提升。结果表明该方法能够有效融合枣品种枣果和叶片器官的特征信息，可为其他品种识别研究提供参考。

Jujube Variety Recognition Method Based on Multi-organ Feature Fusion

Aiming at the problem of jujube variety identification in natural scenes, machine vision technology with jujube fruit as the research object has become one of the mainstream methods for accurate identification of jujube varieties. However, due to the small inter-class difference and large intra-class difference of jujube varieties, it is difficult for a single organ to fully express the different characteristics of jujube varieties. A method of jujube varieties recognition based on multi-organ feature fusion was proposed. Firstly, the YOLO v3 detection algorithm was used to segment and extract the jujube fruit and leaf organs in the collected natural scene images, and a multi-sample dataset of jujube varieties based on Cartesian product was proposed to construct two organ combination pairs, and then based on the EfficientNetV2 network model, a fusion strategy that can fully learn the correlation between the characteristics of the two organs was designed to improve the model performance, and a stepwise transfer training method was introduced to improve the recognition efficiency of jujube varieties. Finally, a large number of experiments were carried out on the constructed dataset containing 20 jujube varieties, and the recognition accuracy of 97.04% was obtained, which was significantly better than that of the existing research results, and the training time and convergence speed of the proposed method were also improved. The results showed that this method can effectively integrate the characteristic information of jujube fruit and leaf organs of jujube cultivars, which can provide valuable reference for other variety identification research.