基于“深度”神经网络的贮藏期柚果品质检测模型研究

Research on quality detection model of pomelo during storage based on“depth”neural network

针对柚果单个理化指标无法表征整果品质,和其内部品质无损检测精度不佳等问题,采用可见-近红外光谱、机器视觉和动态称重等无损检测技术,搭建动态无损检测试验样机,构建“深度”神经网络模型,探究柚果内部综合品质检测方法。研究发现,采集柚果的多特征信息(光谱特征、果形描述子和动态质量等),经数据融合和分析,构建综合品质指标(HP和STP),建立BP神经网络模型,可准确检测蜜柚和沙田柚内部品质,R_(pre)~2分别达到0.930 1和0.936 1,均高于其他内部品质指标(SSC,TA,MC和E)检测结果,具备高效、快速且精度高等优势。研究为厚皮水果综合品质指标构建和模型检测提供参考。

For the problem that a single physical and chemical index of pomelo can not characterize the whole fruit quality and the non-destructive detection accuracy of its internal quality is poor,visible and near-infrared spectroscopy,machine vision and dynamic weighing technology were used to build dynamic non-destructive detection test prototype and build neural network model to investigate the comprehensive quality detection method of pomelo.It was found that by collecting multi-feature information of pomelo(spectral features,fruit shape descriptors,dynamic mass,etc.)and through data fusion and analysis,the comprehensive quality indexes(HP and STP)were built,and a BP neural network model was built,which can accurately detect the internal quality of honey pomelo and Shatian pomelo.Rpre2 reaches 0.9301 and 0.9361,respectively,which are higher than the detection results of other internal quality indexes(SSC,TA,MC and E).It has the advantages of higher efficiency,fast speed and high precision.The results can provide reference for the establishment of comprehensive quality index and model detection of thick-skinned fruits.