本文研究了基于高光谱图像分割的固态发酵中不同杆菌快速识别问题，实验采用马铃薯葡萄糖琼脂培养基对芽孢杆菌、乳杆菌、红螺菌样本进行制备，通过黑色镜头盖在镜头上获取全黑的反射图像，对图像进行去噪、分割处理，获取高光谱图像梯度及分布峰值，利用掩模图像中菌落所处位置对光谱数据信息进行提取，把获取的三种细菌的光谱数据分割成校正集与测试集，依次用于模型构建与验证。通过标准正态变量变化方法完成对原始光谱数据的预处理，利用主成分分析法进行降维处理并区分不同菌落，并采用偏最小二乘判别分析法建立识别模型，分析细菌高光谱图像和高光谱响应值主成分，建立并验证模型。结果表明，采用标准正态变量变化法进行预处理后，高光谱分析总误差与分析误差的比值是1.12，低于规定稳健性的参数；不同细菌菌落的高光谱波峰存在差异，对光的反射率也有较大不同；当培养时长为36 h时，三种细菌可被有效区分；本文模型比其它两种模型更优，对测试集的识别率为98.25%。综上所述，采用本文方法具有分析误差低、识别率高的优点。

In this work, the fast identification of different bacteria in solid-state fermentation based on hyperspectral image segmentation was investigated. The samples of Bacillus, Lactobacillus and Rhodospirillum were prepared by using potato glucose agar medium. The full-black reflective images were obtained on the lens through the black lens cover. The image was denoised and segmented to obtain the gradient and distribution peak value of hyperspectral image. The information of spectral data was extracted from the position of colonies in the mask image, and the spectral data of three bacteria were divided into correction set and test set, which were used to build and validate the model in turn. The pre-processing of original spectral data was completed by the standard normal variable change method. The principal component analysis method was used to reduce the dimension and distinguish different colonies. The identification model was established by partial least squares discriminant analysis. The principal components of hyperspectral image and hyperspectral response value of bacteria were analyzed, and the model was established and validated. The results showed that the ratio of total error to analysis error was 1.12 after pretreatment with standard normal variable change method, which was lower than the prescribed robustness parameter. The hyperspectral peaks of different bacterial colonies were different, and the reflectivity of light was quite different. When the culture time was 36 h, the three bacteria could be distinguished effectively. The model in this work was better than the other two models. The recognition rate of test set was 98.25%. In conclusion, this method had the advantages of low analysis error and high recognition rate.

河南科技厅科技攻关项目（182102210572）