基于小波变换的高光谱散射图像特征提取

Feature extraction of hyperspectral scattering image based on wavelet transformation

高光谱散射图像的特征提取是影响模型精度的重要因素。本文对600个'Golden Delicious'苹果样本的高光谱散射图像进行分析,分别采用平均反射法和小波变换提取特征。小波变换以Danbechies小波系的Db1函数作为基函数进行1层和2层小波分解,然后选取小波低频系数的一范数作为特征值。利用Kennard-Stone算法划分样本,450个样本用于建模,150的样本用于预测。不同方法提取的特征值输入结合偏最小二乘(PLS)算法建立苹果内部品质的预测模型。结果表明1层小波变换特征提取方法与平均反射(mgan reflectance,Mean)特征提取方法相比能将硬度的预测集相关系数从0.797提高到0.821,预测集均方根误差保持不变;糖度的预测集相关系数从0.837略微提高到0.842并降低了预测集均方根误差。因此小波变换为高光谱散射图像提供了一种有效的特征提取方法。

Feature extraction of hyperspectral scattering image is an important factor for model accuracy. This research analyses 600 images of ＇Golden Delicious＇ apple, using mean reflectance method and wavelet transformation to extract feature respectively. Data is decomposed into one layer and two layers with wavelet transformation by using Dbl of Danbechies wavelet series as basis function, then one-norm of the low frequency wavelet coefficients is selected as features. Kennard-Stone algorithm is used to divide samples, 450 samples are used for modeling, another 150 samples are used for prediction. The features of different methods coupled with and partial least squares algorithm （PLS） to develop the prediction model of apple internal qualities. The results show that the feature extraction method based on one layer wavelet transformation which compares with that based on mean reflectance （mean reflectance, Mean） can yield good predictions of fruit firmness with correlation coefficient of prediction set increased from 0.797 to 0.821 and the same root mean square error of prediction set, predictions of fruit soluble solids content with correlation coefficient increased slightly from 0.837 to 0.842 and root mean square error of prediction set reduced. The proposed wavelet transformation method will provide an effective method of feature extraction.