基于注意力机制的葡萄品种多特征分类方法

针对田间自然背景下葡萄品种鉴别缺乏有效识别方法的问题，提出了一种基于融合注意力机制的残差网络ResNet50-SE，对自然背景下不同生长时期的葡萄品种进行分类鉴别，分析并验证了网络的识别效果。将SE注意力模块引入ResNet-50网络，并通过迁移学习实现基于不同时期下葡萄的嫩梢、幼叶及成熟叶片特征的识别；同时为了揭示注意力机制的作用机制，利用Grad-CAM可视化方法，对ResNet50-SE模型每一层所提取的不同生长阶段下的葡萄特征进行可视化解释；通过t-SNE算法对模型提取到的不同葡萄品种的多特征进行聚类分析，进而直观评估模型对多特征提取的性能。结果表明：提出的ResNet50-SE网络在田间复杂背景条件下对于葡萄不同时期的多特征识别具有较高的识别率和较强的鲁棒性，模型测试集准确率达到88.75%，平均召回率达到89.17%，相比于AlexNet 、GoogLeNet、ResNet-50、VGG-16，测试集准确率分别提高了13.61、7.64、0.70、6.53个百分点；注意力机制能明显降低背景影响，强化有效特征；模型对训练集提取的不同生长时期的特征聚类效果较强。可见，SE模块可明显提升ResNet-50模型在特征提取过程的效果，有效降低田间复杂背景对分类结果的影响，为田间复杂背景下葡萄品种的分类识别及田间多特征分类问题提供借鉴。

Multi-features Identification of Grape Cultivars Based on Attention Mechanism

In view of the lack of effective identification methods for grape cultivars identification under the field natural background, a residual network ResNet50-SE based on attention fusion mechanism was proposed to classify and identify grape varieties in different growth periods under natural background, and the identification effect of the network was analyzed and verified. The SE attention module was introduced into ResNet-50 network, and the recognition of grape shoots, young leaves and mature leaves in different periods was realized through transfer learning. Besides, in order to reveal the attention mechanism, the grape characteristics of different growth stages extracted from each layer of ResNet50-SE model were visualized and explained by the Grad-CAM visualization method. The t-SNE algorithm was applied to cluster the multi-features of different grape varieties extracted by the model, and then the performance of multi-features extraction of the model was intuitively evaluated. The results indicated that the ResNet50-SE network had a high recognition rate and strong robustness for grape multi-features recognition in different periods under the complex background conditions in the field. The accuracy rate of the model test set reached 88.75%, and the average recall rate reached 89.17%. Compared with AlexNet, GoogLeNet, ResNet-50 and VGG-16, the accuracy of the test set was improved by 13.61, 7.64, 0.70 and 6.53 percentage points. The attention mechanism can significantly reduce the influence of the background and strengthen the effective features. The model had a strong clustering effect on the features of different growth periods extracted from the training set. Therefore, the SE module can obviously improve the effect of ResNet-50 model in the feature extraction process, and effectively reduce the impact of field complex background on the classification results. The research result can provide a reference for the classification and recognition of grape cultivars multi-features under field complex background.