【摘要】 六安瓜片茶的品质分级长期依赖人工感官评估或电化学分析方法，存在效率低和成本高等局限。针对现有加入注意力机制的神经网络模型在茶叶细粒度分类中的局限性，提出了基于改进嵌入卷积注意力模块（convolutional block attention module, CBAM）的残差网络（residual network, ResNet）模型CBAM-ResNet18的智能评级系统。首先，将CBAM-ResNet18输入层中的最大池化层替换为卷积层，在降低空间维度的同时保留更多茶叶表面微观结构细节。其次，对CBAM-ResNet18的通道数进行压缩，使模型聚焦于不同等级茶叶间差异的关键区域，在提高模型分类精度的同时实现模型轻量化。实验结果表明，在涵盖4个主流品牌、4类等级的自制六安瓜片茶数据集上，所提模型的Top-1准确率达95.0%，媲美现有主流先进注意力网络。而模型总参数量和每秒可执行的浮点运算次数较原CBAM-ResNet18分别下降约83%和30%。在同步开发的低成本便携设备上部署后，采用FP32精度的推理时间约为1.5 s/帧，量化为INT8精度后的推理时间约为0.9 s/帧，证明了系统在六安瓜片茶质检领域的应用潜力。更多还原

【Abstract】 The quality grading of Lu’an Guapian tea has long relied on manual sensory evaluation or electrochemical analysis methods, suffering from limitations of low efficiency and high costs. Aiming at limitations of existing neural network models with attention mechanism in fine-grained tea classification, an intelligent rating system was proposed using an improved CBAM-ResNet18 model based on the residual network(ResNet) model with embedded convolutional block attention module(CBAM). Firstly, the max pooling layer in CBAM-ResNet18 was replaced by a convolutional layer, to retain more tea surface microstructural details during spatial dimension reduction. Secondly, the number of channels of CBAM-ResNet18 was cut, enabling the model to focus on key areas of differences between different grades of tea leaves, while increasing classification accuracy and making the model lightweight.Experimental results show that, on the dataset of self-made Lu’an Guapian tea covering 4 mainstream brands and 4 grades, the model achieves a Top-1 accuracy of 95. 0%, comparable to existing mainstream advanced attention networks. Parameters and floating point operations per second(FLOPs) are reduced by 83% and 30% respectively compared with original CBAM-ResNet18. When deployed on a low-cost portable device developed synchronously, the inference time using FP32 accuracy is approximately 1. 5 s/frame and the inference time quantified to INT8 accuracy is approximately 0. 9 s/frame, validating the application potential of the system in the quality inspection of Lu’an Guapian tea.更多还原