小麦作为主要的粮食作物在我国农业生产、 运输、 食品加工等方面占有重要地位。 不完善籽粒严重影响了小麦质量与粮食安全。 不完善籽粒主要在生产、 存储、 包装等过程中产生, 目前我国小麦质量检测多以人工分选为主, 但存在人主观性较强, 肉眼易疲劳, 且费时费力等问题, 因此, 如何快速准确鉴别小麦不完善粒是现阶段提高生产率和保证粮食安全的重要问题。 运用高光谱成像技术和特征波段选取方法提出一种快速有效的小麦不完善粒鉴别方法。 利用近红外高光谱成像系统获得1 000粒小麦样本在862.9～1 704.2 nm共256个波段的高光谱反射图像, 其中包括健康粒、 生芽粒、 霉变粒和赤霉粒各250粒, 提取每个样本感兴趣区域的平均反射率光谱作为分类特征。 本文首先对提取的全波段光谱信息进行窗口平滑、 一阶导数差分、 矢量归一化等数据预处理, 将原始光谱数据的隐藏信号放大并消除随机误差; 在预处理的基础上运用伪偏最小二乘(DPLS)和正交化线性判别分析(OLDA)对光谱进行特征提取, 降低数据的冗余度; 最后采用仿生模式识别（BPR）建立四类小麦的鉴别模型。 实验结果表明, 采用全波段光谱信息建立的小麦不完善粒鉴别模型的平均识别精度达到97.8%, 分析结果可知, 利用近红外高光谱成像技术的全波段光谱信息对小麦不完善粒鉴别是可行的。 尽管全波段光谱信息取得了较好的鉴别效果, 但高光谱成像设备较为昂贵, 获取高光谱全波段光谱信息数据量较大, 无法满足对现场设备运算速度的高要求, 因此, 采用连续投影算法（SPA）对全波段光谱数据进行特征波段的选择, 使波段数量由256维降低到10维, 从而提高系统的可行性和运算速度。 采用选取的10个特征波段建立小麦不完善粒鉴别模型, 实验结果表明10个特征波段的平均识别精度仅为83.2%, 分析结果可知, 尽管采用10个特征波段提高了系统实时性, 但鉴别准确性较差。 为达到与全波段特征基本相当的鉴别效果, 利用光谱特征与图像特征结合的方法建立小麦不完善粒鉴别模型, 将上述选取的10个特征波段的形态信息、 纹理信息和光谱信息进行结合, 实验结果表明, 10个特征波段的光谱信息与图像信息结合使鉴别的平均识别精度达到94.2%, 此识别效果与利用全波段光谱数据的识别效果基本相当。 利用高光谱成像系统探索了小麦不完善粒鉴别的可行性, 通过分析以上实验可知, 基于近红外高光谱成像技术对小麦不完善粒检测具有良好的效果, 在有效的提高运算速度的同时也保证了系统的鉴别精度, 为后期小麦不完善粒快速检测设备的开发提供了有效的研究方向。

Wheat is a major food crop and occupies an important position in Chinese agricultural production, transportation, and food processing. Unsound kernel seriously affects wheat quality and food security. Wheat unsound kernel is mainly produced during production, storage, and packaging. At present, the manual sorting method is the main method for detecting wheat kernel quality in China. It is subjective, time-consuming, laborious, and costly. Therefore, the rapid and accurate identification method of the wheat unsound kernel will increase productivity and ensure food security. So the method for rapid and accurate detection of wheat unsound kernel was proposed by using hyperspectral image technology and the method of characteristic band selection. In this paper, near-infrared hyperspectral imaging system was used to collected hyperspectral reflection image of 1 000 wheat kernels (including healthy kernels, sprouted kernels, mildewed kernels, and kernels infected with Fusarium head blight, their respective amount are 250) at 862.9~1 704.2 nm (a total of 256 bands) and the average reflectivity of each sample were extracted from region of interests of hyperspectral images as classification characteristics. This paper conducts pre-processing for the extracted full-wave bands spectral information through window smoothing, first order derivative and vector normalization. It will also amplify hidden signals of the original spectral data and erase random errors. On the basis of pre-processing, feature extraction by applying discriminant partial least squares (DPLS) and orthogonal linear discriminant analysis (OLDA) to lower the redundancies of the data. Finally, it establishes identification model for 4 kinds of wheat through pattern recognition (BPR). The experiment results showed that the average identification accuracy of the model for wheat unsound kernel established by using full-wave bands spectral information is 97.8%. The analysis also proves that it is feasible to detect wheat unsound kernel by using near-infrared hyperspectral imaging technology. Though full-wave bands spectral information achieved better detection effect, the high costs of hyperspectral imaging equipment and large amount of hyperspectral full-wave bands spectral information data fail to meet the high requirement of calculation for site equipment. Therefore, this paper uses successive projections algorithm (SPA) to select characteristic bands among full-wave bands data and lower the number of band from 256 dimensions to 10 dimensions to improve the operation and calculation speed of the system. So 10 characteristic bands were taken to establish identification model for wheat unsound kernel. The experiment results showed that the average identification accuracy of the 10 characteristic bands is only 83.2%, which means that though the 10 characteristic bands improve the real-time capability of the system, but they show worse identification accuracy. In order to achieve the identification effect that is basically equivalent to the characteristics of the whole band, this paper uses the combination of spectral features and image features to establish identification model of the wheat unsound kernel. All kinds of relevant information (morphological information, texture information, spectral information) of the kernel of the above 10 selected wave bands are integrated. The experimental results showed that the combination of spectral information and image information in 10 characteristic bands can achieve an average identification accuracy of 94.2%. Its identification effect is basically consistent with the use of full-wave bands spectral data. This paper uses hyperspectral imaging system to explore the feasibility of wheat unsound kernel detection. From the analysis of the above experiment, it can be seen that near-infrared hyperspectral imaging technology shows better results in the detection of wheat unsound kernel. It can guarantee the identification accuracy of the system while improving calculation speed so it offers an effective research orientation for later development of equipment that is able to detect wheat unsound kernel in a quick way.