红提糖度是重要的内部品质衡量指标, 传统的检测方法均为破坏性生化检测, 本文基于高光谱成像技术, 提出了一种基于高光谱信息融合的红提糖度含量无损检测方法。采集并提取260个红提样本的光谱信息和图像信息, 对光谱信息分别利用SNV、S-G等光谱预处理方法建立PLSR模型, 确定最好的光谱预处理方法, 分别采用一次降维(GA、CARS、IRIV)算法和组合降维算法(CARS-SPA、IRIV-SPA、GA-SPA)共六种降维方法对光谱信息进行特征变量提取; 获取灰度共生矩阵的图像纹理信息, 结合图像的颜色信息(R、G、B、H、S、V、L、a、b), 组成19个图像特征参数, 采用PCA算法对图像信息进行降维, 分别建立基于降维处理后的光谱信息、图像信息以及两者融合的红提糖度线性预测模型PLSR、非线性预测模型LSSVM, 并对比分析模型的优劣。结果表明, 若只利用光谱信息建模, IRIV-SPA可有效地提取红提糖度光谱信息的特征波长, 提高模型的预测性能; 若只利用图像信息进行建模, 模型的预测性能不好, PCA降维有效地提高了模型的预测性能, 但提高的性能有限; 将IRIV-SPA特征波段提取后的光谱和经PCA降维后的图像信息进行融合, 分别建立PLSR和LSSVM模型, 红提糖度的最优PLSR模型的校正集和预测集相关系数分别为0.943, 0.941; 红提糖度的最优LSSVM模型的校正集和预测集相关系数分别为0.954, 0.952。LSSVM所建模型的效果好于PLSR所建模型, 但模型的运算时间较长。两种模型的精度均比单方面基于光谱或图像信息的模型都有较大的提高, 表明融合高光谱图像的光谱与图像信息不仅可以提高模型的运算速度、简化模型, 同时有效地提高了红提糖度预测性能, 为红提糖度的检测找到了一种新的方法。

The sugar content of red grapes is an important index to measure its internal quality. The conventional methods based on biochemistry to test the sugar content are all destructive. This article proposed a method using high spectrum technology for sugar content test which does not break the fruit at all. The spectral and graphical information is collected and extracted from 260 samples. PLSR model was set up by the usage of preprocessing method like SNV and S-G, which is aiming to find out the fittest one. We adopted six algorithms including GA, CARS and IRIV from one-time dimension-reducting algorithm and CARS-SPA, IRIV-SPA and GA-SPA from combined dimension-reducting algorithm respectively, to extract the characteristic variable from the spectral information. 19 characteristic parameters were obtained from the texture information of images by gray-level co-occurrence matrix and the color information respectively by RGB model. After reducing the dimensions of image information with PCA algorithm, we built the linear predictive model PLSR and non-linear predictive model LSSVM, which were based on the dimension-reduced spectral information, image information and the combination of the formers separately, then compare the efficiency of models. The results show that, when we modeled with spectral information merely, the IRIV-SPA was able to extract the characteristic wavelength of the spectrum information of red grapeｓ sugar content; comparely, the model built with information of images only did not work well, even though the PCA that reduced the dimensions effectively increased the performance of modelｓ prediction but limitedly. We built the PLSR model and LSSVM model with both spectral wave-range characteristics by IRIV-SPA and image information dimension-reduced by PCA. The correlation coefficient of optimal PLSR modelｓ calibration set and prediction set is 0.943 and 0.941 respectively while 0.954 and 0.952 for LSSVM model. The performance of model of LSSVM was better than PLSR, but the former one took more time for operation. The relatively greater promotion on accuracy of both models based on the combination of two classifications of characteristics indicated that the combination of spectral information and image information was not just able to increase the operating speed and simplify the model but effectively promote the performance of the prediction for red grapeｓ sugar content. Thus a new method for the detection of red globe grape sugar content was found.