Study on Fruit Brix by Hyperspectral Detection Based on Machine Learning

• Main Authors: WU, BO-HUEI, 吳柏輝
• Other Authors: Chang, Ching-Lung
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/j577qk

碩士 === 國立雲林科技大學 === 資訊工程系 === 107 === In this paper, the Vis-NIR hyperspectral imaging system is used to predict the sweetness of fruit, in order to explore the effective screening of hyperspectral data features to improve the overall fruit sweetness prediction performance. Taking the sweetness of "Carambola" as the research index, the hyperspectral imaging technique was used to collect the reflectivity images of each carambola at 400~1000nm, and the samples were preprocessed by image segmentation. Various heuristic algorithms were used to select hyperspectral wavelengths. Remove unnecessary data noise. Based on different screening methods, a partial least squares method (PLS) regression prediction model was established, and the prediction coefficients were evaluated by the coefficient of determination (R^2-score) and root-mean-square error (RMSE). Feedback back to the heuristic algorithm to find a better combination of hyperspectral wavelength predictions. In addition, for the topic of "continuous variable prediction", we try to use a neural network to estimate the possibility of combining hyperspectral imaging technology with a neural network.