Abstract: A multistructure neural network (MSNN) was proposed and applied to classify five classes of rice grains. The MSNN model consisted of five parallel multilayer feedforward neural networks (MLNN). With two hidden layers MLNN was trained using morphological and color features of the rice grains extracted from their images as input. The average classification accuracy of MSNN was 92.66%, with an increase of over 504 percent points than that of MLNN; moreover the network training time for MSNN was shorter than that for MLNN.

Key words: neural network, recognition, rice grain, morphological feature, color feature, appearance quality, image processing

摘要： 应用多结构神经网络建立了大米外观品质评判模型，可实现5类大米的识别。模型采用5个并行工作的多层前向神经网络。每个多层前向神经网络包含两个隐含层，以大米图像的形状特征和颜色特征作为网络输入。网络训练和仿真结果显示模型识别的平均准确率为92.66%，比相同网络复杂度下的多层前向神经网络模型提高5.04个百分点，并且网络学习速率快。

关键词: 神经网络, 识别, 大米, 形状特征, 颜色特征, 外观品质, 图像处理