基于卷积神经网络的鲜茶叶分类方法研究

Research on Classification Method of Fresh Tea Leaves Based on Convolutional Neural Network

摘要: 针对鲜茶叶分类任务中的技术瓶颈，本文提出一种基于AlexNet的改进架构，通过四项核心技术有效突破传统方法的局限性。首先，在AlexNet模型Conv3~Conv5层后嵌入通道注意力机制(Squeeze-and-Excitation, SE)模块，自适应地调整通道的特征响应以增强模型对特征的表达能力；其次，在全卷积层后添加批归一化(Batch Normalization)层，使鲜叶数据集的训练收敛轮次由50轮优化至20轮；另外，在模型输出层前加入全局平均池化(Global Average Pooling, GAP)层，成功将模型参数量从178.39 M压缩至14.3 M。实验结果表明，所提出的改进方案在保证分类精度的前提下，显著提升了模型的训练效率与泛化能力。

Abstract: In order to solve the technical bottleneck in the task of sorting fresh tea leaves, this paper proposes an improved architecture based on AlexNet, which effectively breaks through the limitations of traditional methods through four core technologies. Firstly, the Squeeze-and-Excitation (SE) module was embedded in the Conv3~Conv5 layers of the AlexNet model to adaptively adjust the feature response of the channel to enhance the model’s ability to express features. Secondly, a Batch Normalization layer was added after the full convolutional layer, so that the training convergence rounds of the fresh leaf dataset were optimized from 50 rounds to 20 rounds. In addition, the Global Average Pooling (GAP) layer was added before the model output layer, and the model parameters were successfully compressed from 178.39 M to 14.3 M. Experimental results show that the proposed improved scheme significantly improves the training efficiency and generalization ability of the model under the premise of ensuring classification accuracy.