Método para classificação de sementes agrícolas em imagens obtidas por tomografia de raios-X em alta resolução

• Autor(a) principal: Brito, André Ribeiro de
• Data de Publicação: 2020
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFSCAR
• Texto Completo: https://repositorio.ufscar.br/handle/ufscar/13236
• topic: Processamento digital de imagens; Tomografia de raios-X; Classificadores supervisionados; Classificação de sementes agrícolas; Digital Image processing; X-ray tomography; Supervised classifiers; Agricultural oilseeds; CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO; CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO

This research produced a method for classifying agricultural oil seeds based on tomography (CT) assays. Oilseeds may present good or poor quality for planting, the latter being related to the presence of defects associated with physical imperfections such as cracks, breaks and voids. In such a context, for the development of the method, high-resolution tomographic slices reconstruction have been considered, as well as the evaluation of digital image processing and visualization techniques, i.e., evaluations related to pre-processing, segmentation, extraction of supervised characteristics and classifiers. For such evaluations, samples containing oilseeds with and without defects for sunflower (Helianthus annuus L.), physic nut (Jatropha curcas L.), and soybean (Glycine max ( L.) Merrill) were classified. For the choice of the pre-processing, processing and visualization of tomographic images techniques, the following metrics were used: PSNR (Peak Signal-to-Noise Ratio), MSE (Mean Square Error), SSIM (Structural Similarity Index Measure), and DSC (Dice Similarity Coefficient). Thus, for the establishment of the method, the techniques for filtering and segmentation have been selected considering the Nonlocal means and Gaussian filters, while for the segmentation steps, the simple thresholding and graph techniques have been selected, respectively. For the characteristic extraction stage, the HOG (Histogram of Oriented Gradients) and the Hu invariant moments techniques have been selected, as they can allow obtaining texture descriptors and descriptors based on geometric characteristics, respectively. Additionally, the PCA (Principal Component Analysis) technique has been used to establish a composite characteristic vector with these descriptors. Finally, not only the NB (Naïve Bayes) but also the SVM (Support Vector Machine) classifier shown its usefulness. Furthermore, for the method, the NB classifier has been indicated for the analysis of tomographic slices of sunflower seeds and the SVM classifier for the analysis of tomographic slices of jatropha and soybean seeds. The method developed for the grading of oilseeds has been shown to be suitable for selection in pre-planting, which helps to assist decision making in the selection of good quality agricultural seed. It also contributes to the evolution of advanced techniques and analysis tools, obtained from digital image processing and visualization techniques.