基于深度主动学习与CBAM的细粒度菊花表型识别

针对菊花种类繁多，花型差别细微，准确标注比较困难的问题，基于深度主动学习与混合注意力机制模块(Convolutional block attention module,CBAM)，提出了一种标号数据不足情况下的菊花表型智能识别方法和框架。首先，通过主动学习策略基于最优标号和次优标号法（Best vs second best,BvSB）在未标记菊花样本中选取信息量较大的样本进行标记，并将标记后的样本放入训练样本中；其次，使用深度卷积神经网络ResNet50作为本文的主干网络训练标记样本，引入混合注意力机制模块CBAM，使模型能够更为准确地提取细粒度图像中的高层语义信息；最后，用更新后的训练样本继续训练分类模型，直到模型达到迭代次数后停止。实验结果表明，该方法在少量菊花标记样本下，精确率、召回率和F1值分别达到93.66%、93.15%和93.41%。本文方法可为标号数据不足情况下的菊花等花卉智能化识别提供技术支撑。

Fine-grained Chrysanthemum Phenotype Recognition Based on Deep Active Learning and CBAM

Chrysanthemums have a wide variety of flower types with subtle differences in flower phenotypes, which are difficult to label accurately, and this poses a great challenge for intelligent classification and recognition of chrysanthemums. Based on deep active learning and hybrid attention mechanism module, i.e. convolutional block attention module (CBAM), a method and framework for intelligent recognition of chrysanthemum phenotypes under insufficient labeling data was proposed. Firstly, the more informative samples among the unlabeled chrysanthemum samples were selected for labeling by an active learning strategy based on the optimal labeling and second-optimal labeling method BvSB (Best vs second-best), and the labeled samples were put into the training samples;secondly, a deep convolutional neural network ResNet50 was used as the backbone network to train the labeled samples, and the hybrid attention mechanism module CBAM was introducted, so that the model can more accurately extract the high-level semantic information in fine-grained images;finally, the classification model continued to be trained with the updated training samples until the model reached the number of iterations and then stopped. The experimental results showed that the method can achieve 93.66%, 93.15% and 93.41% of precision, recall and F1 value respectively with a small number of chrysanthemum labeled samples. The method can provide technical support for intelligent identification of chrysanthemums and other flowers under the situation of insufficient labeling data.