Fusão de sensores de baixo custo aplicada à seleção e classificação automatizada de mangas para exportação.

Abstract

Agricultural products, especially fruits and vegetables, are subject to physical and physiological damage during all production steps: planting, growing, harvesting, transport to packing houses, washing and cleaning, sorting and classification, packaging and transportation to the final consumer. A key factor to ensuring competitiveness in globalized markets is the ability of a given production chain to provide high quality products, with low cost and suitable for the final consumer, while reducing costs, losses and damage to the lowest possible levels, compatible with those obtained by competitive producers. In this context, many studies have been carried out worldwide, aiming at the development and application of new technologies, including sensors and automation systems, for quality assessment, selection and classification of fruits and vegetables in general. Significant progress has been achieved in the evaluation of internal and external qualitative parameters, especially at laboratory level. However, real automation of postharvest processing lines, at low cost, is still a challenge, due to the complexity of sorting: fruits and vegetables have variability inherently greater than the equivalent processed products, especially as a result of their own particular characteristics during the production stages. All of these stages, from planting through growth, harvest and post-harvest, including climate, pests and diseases, affect directly the variability and the final quality of the product, increasing investments in automation for a proper selection and classification with focus on the final consumer. In addition to research aimed at automation of these processes, it is necessary to develop grading equipment by volume / weight with national (Brazilian) technology. Such equipments should be more accessible, have low cost and present lower operation and maintenance costs, so they can be used by small farmer cooperatives. In Chapter 1, the morphological characteristics and the main defects in Tommy Atkins mangoes were identified. In Chapter 2, the goal was to assess the dimensional parameters of the fruit, using b&w images and low cost video cameras. In Chapter 3, the work aimed to estimate the mass of mangoes using color images captured with a commercial digital camera. The objective of Chapter 4 was to use different mechanical sensor technologies, of low cost, simultaneous and integrated ones, for fruit firmness assessment. The purpose of Chapter 5 was to develop a methodology for sensor fusion for a better estimate of non-destructive firmness, demonstrating this technique feasibility for the determination of a maturity index by means of parameters extracted from data of three different sensors, a video camera and two acceleration sensors, all at low cost. Artificial neural networks were used for sensor fusion. Final considerations describe the main contributions of this work: modification in the Matlab® standard algorithm (Otsu's method) for image processing, in order to improve the segmentation of fruit images; identification of the projected area parameter as the best estimator of the mass of Tommy Atkins mangoes, through captured image processing with low cost cameras; proposition of a new combined reference parameter for estimating the firmness of Tommy Atkins mangoes, when using the low-mass impact technique; demonstration of the feasibility of MEMS accelerometer use and a sealed microphone electret, both of low cost, to estimate the firmness of Tommy Atkins mangoes, when using the low-mass impact technique; use of the Shock Response Spectrum of a MEMS accelerometer and the signal amplitude (peak) of a sealed electret microphone as non-destructive firmness estimators of Tommy Atkins mangoes, when using the low-mass impact technique; use of percussion point for positioning the acceleration sensors when using the low-mass impact technique; development of a level conversion board (FACS1), of low cost, for direct interfacing between acceleration sensors and computer sound cards; use of artificial neural networks, as sensor fusion technique, improved significantly (about 19.5%) the correlations of nondestructive firmness estimators in relation to the reference parameters, in comparison with the estimators used in separate.