Purpose

This paper proposes a novel deep learning based method towards the identification of a pistachio tree cultivar from its image.

Design/methodology/approach

The investigated scope of this study includes Iranian commercial pistachios (Jumbo, Long, Round and Super long) trees. Effective use of high-resolution images with standard deep models is addressed in this study. A novel image patches extraction method is also used to boost the number of samples and dataset augmentation. In the proposed method, handcrafted ORB features are used to detect and extract patches which may contain identifiable information. An innovative algorithm is proposed for searching and extracting these patches. After extracting patches from initial images, a Convolutional Neural Network, named EfficientNet-B1, was fine-tuned on it. In the testing phase, several patches were extracted from the prompted image using the ORB-based method, and the results of their prediction were consolidated. In this method, patch prediction scores were in descending order, sorted by the highest score in a list, and finally, the average of a few list tops was calculated and the final decision was made.

Findings

Examining the proposed method on the test images led to an achievement of a recognition rate of 97.2% accuracy. Investigation of decision-making in the test dataset could reveal that this method outperformed human experts.

Originality/value

Cultivar identification using deep learning methods, due to their high recognition speed, lack of specialist requirement, and independence from human decision-making error is considered as a breakthrough in horticultural science. Variety cultivars of pistachio trees possess variant characteristics or traits, accordingly recognising cultivars is crucial to reduce the costs, prevent damages and harvest the optimal yields.

中文翻译：

---

阿月浑子树品种识别的有效深度学习模型

目的

本文提出了一种基于深度学习的新颖方法，用于从开心果树图像中识别开心果树。

设计/方法/方法

本研究的研究范围包括伊朗商业开心果树（巨型，长树，圆形和超长树）。本研究解决了将高分辨率图像与标准深度模型一起有效使用的问题。一种新颖的图像补丁提取方法也用于增加样本数量和数据集扩充。在提出的方法中，手工制作的ORB功能用于检测和提取可能包含可识别信息的补丁。提出了一种创新的算法来搜索和提取这些补丁。从初始图像中提取补丁后，对其进行了微调，即名为EfficientNet-B1的卷积神经网络。在测试阶段，使用基于ORB的方法从提示的图像中提取了几个补丁，并对它们的预测结果进行了合并。用这种方法

发现

在测试图像上检查所提出的方法可以实现97.2％的识别率。对测试数据集中的决策进行调查可以发现，该方法的性能优于人类专家。

创意/价值

由于深度学习方法的识别速度快，缺乏专家要求以及不受人为决策错误的影响，因此使用深度学习方法进行的品种识别被认为是园艺科学领域的一项突破。阿月浑子树的品种具有不同的特征或性状，因此，认识到品种对于降低成本，防止损害和收获最佳产量至关重要。