基于高光谱图像与视觉词袋模型的稻种发芽率预测研究

Prediction of Rice Seed Germination Rate Based on Hyperspectral Image and Bag of Visual Words Model

为实现稻种品质的快速鉴定,以稻种最重要的品质参数之一——发芽率作为主要评价指标,通过高光谱成像技术结合视觉词袋(Bo VW)模型的方法进行稻种发芽率的分级评价。挑选Y两优302、两优108和内5优8015三个品种的杂交水稻种子各100粒,在温度40℃、相对湿度100%条件下对三种稻种分别老化处理0、1、2、3、4 d,得到5个活力梯度的稻种。采集300粒稻种的高光谱图像,随机分为训练集(200份)和测试集(100份)。图像采集完毕后,进行稻种发芽实验,第14天时计算发芽率。采用主成分分析(PCA)方法选取特征波长,利用密集尺度不变特征变换(SIFT)算法提取稻种图像局部特征,再根据K-means算法聚类生成视觉词典。利用以径向基(RBF)核为核函数的支持向量机(SVM)分类器建立稻种发芽率分级预测模型,判别精度为95.65%。结果表明,采用高光谱成像技术结合视觉词袋模型进行水稻发芽率的快速、无损预测是可行的。

Germination rate is one of the most important quality parameters of rice seeds. In order to identify the quality of rice seeds rapidly, the hyperspectral imaging technology and the bag of visual words （BoVW） are combined to establish a grading model of rice seed germination rate. Three kinds of hybrid rice seeds, YLiangyou302, Liangyou 108 and Nei5you8015 are selected to be aged artificially for 0, 1, 2, 3, 4 d under the condition of temperature of 40 ℃ and relative humidity of 100% , and 5 dynamic gradients are obtained. Hyperspectral images of 300 samples are randomly divided into a training set （200 samples） and a test set （100 samples）. After imaging selection, the germination rate test is performed and the germination rate is calculated on the 14th day. Principal component analysis （PCA） is applied to select characteristic wavelengths from the full spectral band. Scale-invariant feature transform （SIFT） is used to extract the local features of each image. All local features are clustered by K-means algorithm to generate visual dictionary. The support vector machine （SVM） classification model of rice seed germination rate is established with the radial basis function （RBF）, and the discrimination accuracy reaches 95.65%. The result suggests that it is feasible to predict germination rate of rice seeds by using hyperspectral imaging technology combined with BoVM model.