【摘要】 快速准确地识别大米加工精度，是实现大米碾削加工自动控制的关键。采集不同碾削时间的大米图像，通过滤波及压缩等处理后，制成不同精度等级的数据集，再以经典卷积神经网络(Convolutional neural network, CNN)LeNet-5、AlexNet和VGG-16为基础，分别采取了一系列针对性的优化策略，对模型进行改进和优化，得到3种改进的大米加工精度识别模型；通过试验比较了这些改进优化方法对模型性能的影响。结果表明，在引入卷积核尺寸为9×9和11×11的两个子卷积网络分支后，LeNet-5模型识别准确率由82.23%提高到92.53%;在对AlexNet模型去除两个全连接层，并采用普通池化后，模型识别准确率由92.89%提高到96.73%;如果将Dropout层和组归一化(Group normalization, GN)层引入到VGG-16模型，并采取最大相关-最小冗余(Maximum relevance minimum redundancy, MRMR)特征选择算法对模型进行改进优化，模型准确率较原模型提升7.56个百分点，达到98.87%,为所有测试模型中最高。这表明改进的VGG-16网络模型，可以用于对大米加工精度等级进行精准识别。更多还原

【Abstract】 Rapid and accurate identification of the rice processing precision is essential for achieving automated control of rice milling process. Images of rice milled for different durations were collected and processed through filtering and compression to create datasets of various precision levels. Based on three classical convolutional neural network(CNN) architectures, including LeNet-5, AlexNet, and VGG-16, targeted optimization strategies were applied to develop three improved models for rice processing precision identification. Experimental analyses were conducted to compare the performance of these improved methods. The experimental results revealed that introducing two sub-convolutional network branches with kernel sizes of 9×9 and 11×11 to the LeNet-5 model led to the increase in recognition accuracy from 82.23% to 92.53%. Regarding the AlexNet model, removing two fully connected layers and adopting conventional pooling increased its recognition accuracy from 92.89% to 96.73%. Most notably, the VGG-16 model, after incorporating maximum relevance minimum redundancy(MRMR) feature selection, Dropout, and group normalization(GN) layers, achieved an accuracy of 98.87%. This represented a improvement of 7.56 percentage points over the original model, ultimately making improved VGG-16 the most accurate model among all tested models. Consequently, this demonstrated that the improved VGG-16 model was highly effective for accurate identification of rice processing precision levels.更多还原