高光谱图像技术是在种子识别领域广泛应用的农产品品质无损检测方法。 特征信息的充分提取和最优波段的选择是影响高光谱图像技术种子鉴选在线应用的关键因素。 目的在于利用联合偏度算法选择高光谱图像的最优波段, 用于开发在线的种子分级系统。 论文利用高光谱图像采集系统获取10类共960粒玉米种子在438～1 000 nm（共219个波段）波段范围内的高光谱图像, 并提取了种子高光谱图像的平均光谱、 图像熵特征。 利用联合偏度算法选择了高光谱图像的最优波段, 分别建立了基于平均光谱、 图像熵、 平均光谱和图像熵联合特征条件下的支持向量机种子分类模型, 比较不同特征下分类模型的识别精度。 实验结果表明: 无论是全波段分类模型, 还是建立在最优波段基础上的分类模型, 利用平均光谱和图像熵联合特征获得的分类精度均高于平均光谱和图像熵两种单一特征模型。 在10个最优波段条件下, 联合特征分类模型的识别精度达到了96.28%, 比光谱均值和图像熵的识别精度分别提高了4.30%和20.38%, 也高于全波段联合特征识别模型的93.47%。 利用联合特征建立玉米种子分类模型时, 基于联合偏度的波段选择算法的分类精度要高于无信息变量消除法、 连续投影算法和竞争性自适应重加权算法。 该研究为种子高光谱图像识别技术的在线运用提供了可行的途径。

As an effective method for the nondestructive measurement of agricultural products quality, hyperspectral imaging technology has been widely studied in the field of seed classification and identification. Feature extraction and optimal wavelength selection are the two critical issues affecting the application of hyperspectral image in the field of seed identification. This study aimed to select optimal wavelengths from hyperspectral image data using joint skewness algorithm, so that they can be deployed in multispectral imaging-based inspection system for the automatic classification of maize seed. The hyperspectral images covering the wavelength range of 438～1 000 nm were acquired for 960 maize seeds including 10 varieties. After extracting the mean spectrum and entropy from the hyperspectral images, the joint skewness algorithm was used to select optimal wavelengths, and the classification models based on support vector machine were developed using the mean spectrum, entropy, and their combination, respectively. The experimental results indicated that the classification accuracy of the models developed by combination of the mean spectrum and entropy were higher than that of the mean spectrum or entropy for either full wavelengths or optimal wavelengths. The classification model for the combination of the mean spectrum and entropy based on the 10 optimal wavelengths selected by the joint skewness algorithm obtained 96.28% accuracy for test samples, with improvements of 4.30% and 20.38% over that of the mean spectrum and entropy, respectively, which was higher than the classification accuracy of the model that developed in the full wavelength (i.e., 93.47%). Meanwhile, the classification model based on joint skewness algorithm yielded the better classification accuracy than that of uninformative viable elimination algorithm, successive projections algorithm, and competitive adaptive reweighed sampling algorithm. This study made the online application of the hyperspectral image technology available for seed identification.