近红外光谱（NIRS）技术在作物种子品种鉴别上具有一定的可行性, 但如果待测种子的存储时间不同, 识别模型的准确性会受到影响。 为了降低存储时间对识别模型的影响、 提高模型的预测能力, 将NIRS技术与图像处理技术相融合, 提取出与品种生理生化指标相关的光谱特征和与品种相关的表观图像特征。 为了提取出最优的光谱特征, 首先提出一种改进的后向间隔偏最小二乘（IM_BiPLS）光谱区间选择算法。 针对BiPLS分段数难以确定的问题, 让分段数在一定范围内变化, 以每个分段数所取得的组合区间建立模型的相关系数和交叉验证均方根误差之比作为评价指标, 该指标最大时的分段数所对应的波段组合为最优。 然后使用竞争自适应重加权法（CARS）去除IM_BiPLS所选波段中的无信息变量和共线性变量实现光谱特征优选。 为了提取与品种相关的表观图像特征, 首先使用基于最大熵和双重区域标记的图像分割算法完成不感兴趣区域去除和单粒种子图像分割; 然后提取单粒种子的形态、 纹理和颜色特征并计算出每个图像样本所有种子的统计平均特征。 最后使用CARS对这些特征进行深层次优选完成图像特征提取。 以10个黄色玉米品种为研究对象, 采集216个样本的NIRS数据和对应的图像。 针对光谱数据, 使用IM_BiPLS算法从全谱1 845个变量中选出了具有736个变量的波段组合, 使用CARS进一步从中优选出光谱变量29个。 针对图像数据, 提取出图像特征29个, 使用CARS进一步优选出图像特征11个。 分别以IM_BiPLS提取的光谱特征波段、 IM_BiPLS-CARS优选的特征波长、 图像特征(Image)、 CARS提取的图像特征(Image-CARS)以及IM_BiPLS-CARS优选的特征波长融合CARS提取的图像特征（Compound）为输入, 以样本对应的类别为输出, 建立BP神经网络模型。 测试结果表明Compound-BP模型的性能最佳, 训练准确率和验证准确率均为100%, 测试准确率为97.7%。 实验结果说明NIRS特征融合图像特征可以有效地提高识别模型的精度, 降低存储时间对模型的影响, 为实现玉米种子品种的无损、 快速、 精确识别提供参考。

Near-infrared spectroscopy (NIRS) technology has certain feasibility in identifying crop seed varieties, but if the storage time of the seeds to be tested is different, the accuracy of the identification model will be affected. In order to reduce the influence of storage time on the recognition model and improve the models prediction ability, NIRS technology and image processing technology are combined to extract spectral features related to physiological and biochemical indicators of varieties and apparent image features related to varieties. In order to extract the optimal spectral features, an improved backward interval partial least squares (IM_BiPLS) spectral interval selection algorithm is proposed. Aiming at the problem that it is difficult to determine the number of segments of BiPLS, the algorithm changes the number of segments within a certain range and takes the ratio of the correlation coefficient of the model established by the combination interval obtained by each segment number and the root mean square error of cross-validation as the evaluation index. When the index is maximum, the band combination corresponding to the segment number is the best. The competitive adaptive reweighting method (CARS) removes the uninformative and collinear variables in the selected band of IM_BiPLS and further optimises the spectral features. In order to extract the apparent image features related to varieties, firstly, the image segmentation algorithm based on maximum entropy and double region marking is used to remove the regions of interest and segment the single seed image. Then a single seeds morphological, texture and color features are extracted, and the statistical average features of all seeds in each image sample are calculated. Finally, CARS are used to optimize these features to complete image feature extraction. Taking 10 yellow maize varieties as the research object, NIRS data and corresponding images of 216 samples were collected. For spectral data, use IM_BiPLS algorithm selects the band combination with 736 variables from 1845 variables in the full spectrum and uses CARS to optimize 29 spectral variables further; For image data, 29 image features are extracted, and 11 image features are further optimized by CARS. Respectively using the spectral feature band extracted by IM_BiPLS, the preferred feature wavelength extracted by IM_BiPLS_CARS, the image feature(image), the image feature extracted by CARS(image_CARS), and the fusion between IM_BiPLS_CARS and image_CARS(compound) as the input and the corresponding category of the sample as the output to set up BP neural network models. The test results show that the performance of the compound BP model is the best, the training accuracy and verification accuracy are 100%, and the test accuracy is 97.7%. The experimental results demonstrate that the fusion of NIRS features with image features can effectively improve the accuracy of the recognition model and reduce the impact of storage time on the model. This provide a reference method for achieving the non-destructive, rapid and accurate recognition of corn seed varieties.