Name: HATYLLA BRUNO MILAGRE
Publication date: 25/07/2023
Examining board:
Name |
Role |
|---|---|
| ALEX ARBEY LOPERA SEPÚLVEDA | Examinador Externo |
| CARLOS AUGUSTO CARDOSO PASSOS | Presidente |
| MARCOS TADEU D AZEREDO ORLANDO | Examinador Interno |
Summary: Processing and Analysis of Digital Images (PADI) is a set of techniques that alter the pixels of a given image through mathematical operations. The main objective is to modify the image in order to allow better visualization or preparation for analysis on a computer with well-defined routines. The reliability of the results obtained through PADI strongly depends on how image segmentation is performed in the process. In this work, an idealized routine for image segmentation was defined, and through this routine, the results of porosity and apparent density of a SmBa2Cu3O7-/Al composite were obtained. Four different samples of the composite "(1-x)SmBa2Cu3O7- + xAl" were produced with mass fractions where x = 0%, 40%, 50%, and 60%. The samples were sanded and polished for characterization by Scanning Electron Microscopy (SEM). The obtained images were analyzed using a machine learning model for segmentation. In this analysis, the variation of four main parameters was considered: the accuracy in collecting pixel samples to define a region, the number of corrective samples, image equalization in pre-processing, and the classifiers reliability based on the convergence among their results. The results demonstrated that the accuracy in collecting samples requires careful attention from the operator to avoid the intersection of desired regions, that is, overlapping of regions of interest. It is important to emphasize that the samples must contain a sufficient number of pixels with different tonalities for effective pattern recognition by the software. It was also observed that two or more corrective samples compromise the segmentation result. Image equalization proved to be essential for obtaining reliable results and allows for flexibility in the brightness quality requirement of the original images. The results also suggest that checking the reliability of the classifiers based on the convergence among their results contributes to a better segmentation.
