The evolution of the Deep Convolutional Neural Networks (DCNN) has progressively increased their ability to transfer the weights learned with large generic datasets to tasks with smaller collections or more specific data. However, the adjustment of these networks for different domains usually demand a fine-tuning step for which data may not be abundant enough. That is the case of plant species recognition task, which also suffers from class imbalance. Moreover, there is still a large variety of classification effectiveness with the models trained with the features extracted with different networks. All these factors create a complex assessment scenario and demand costly experimental validation procedures. Hence, in the context of plant species recognition, this work performs a comparative study of multiple pre-trained DCNNs to extract deep features from images of multi-organ plant observations. Beyond it, Softmax and six variations of the Support Vector Machine (SVM) classifier were used for the assessment of the suitability of the evaluated DCNNs. The experimental validation demonstrates great effectiveness variances of different DCNNs for feature extraction and the importance of such an experimental assessment for classification accuracy maximization. Beyond it, our results also show that exploiting deep feature extraction and an SVM-based classification outperformed a traditional setting based on neural classifiers. In fact, considering a hyperparameter optimization, the top performing SVM configuration allowed 82% of Micro-F1 in contrast to 76% of the second best (Softmax). The experiments also highlight such behavior with an effectiveness evaluation which specially accounts for dataset imbalance, a usual scenario in plant species recognition.

深度卷积神经网络 (DCNN) 的发展逐渐提高了它们将使用大型通用数据集学习的权重转移到具有较小集合或更具体数据的任务的能力。然而，针对不同领域的这些网络的调整通常需要一个微调步骤，因此数据可能不够丰富。这就是植物物种识别任务的情况，它也存在类别不平衡的问题。此外，使用不同网络提取的特征训练的模型仍然存在多种分类有效性。所有这些因素造成了一个复杂的评估方案，并需要昂贵的实验验证程序。因此，在植物物种识别的背景下，这项工作对多个预训练的 DCNN 进行了比较研究，以从多器官植物观察的图像中提取深层特征。除此之外，Softmax 和支持向量机 (SVM) 分类器的六个变体用于评估所评估 DCNN 的适用性。实验验证表明，不同 DCNN 对特征提取的有效性差异很大，以及这种实验评估对分类精度最大化的重要性。除此之外，我们的结果还表明，利用深度特征提取和基于 SVM 的分类优于基于神经分类器的传统设置。事实上，考虑到超参数优化，性能最好的 SVM 配置允许 Micro-F1 的 82%，而第二好的（Softmax）只有 76%。