Multi sensor system for classifying Harumanis mango based on external and internal quality

Abstract

This thesis presents a multi sensor system for classifying Harumanis mango based on its external and internal quality. Both external and internal quality of Harumanis mango affects the consumer buying preferences. Current method of classifying Harumanis mango is done manually and destructive for its internal quality determination. The proposed system consists of two parts. First part is the external quality classification using machine vision system which is based on its shape and mass. The second part is the internal quality classification using near infrared (NIR) spectroscopy, based on its total soluble solid (TSS) value. An image acquisition platform was built to capture the 3- Dimensional image of Harumanis mango in a single acquisition. A real-time measurement calibration technique was developed in this research. Combination of Fourier descriptor parameters and size-shape parameters was used to recognize the shape of Harumanis mango. An improved two-dimensional disk method was used to estimate the volume of Harumanis mango based on the captures image. Then a correlation between the actual volume and actual mass was derived and used to estimate the mass of Harumanis mango on inline system. The proposed method can correctly classify the Harumanis mango according to its shape and mass 94.2% of the time. NIR spectrometer was used to obtain the reflectance wavelength of the Harumanis mango. The juice from the mango was obtained and measured with a refractrometer to obtain the actual TSS value. Then, the acquired NIR wavelength was analysed and correlate with the actual TSS value using multivariate analysis. A regression value of 0.85 for calibration set was found from the analysis, which explained that there was a high correlation between the wavelength and TSS. Stepwise Discriminant analysis method was used to find the significant wavelength that can be used to determine the maturity stage in real-time system. Ten wavelength points were selected and verified on the testing set. The discriminant model can be accurately determined the maturity stage with 85.0% accuracy