Segmentação de nanotubos de carbono em imagens de microscopia eletrônica de transmissão

Abstract

Nowadays, digital image processing is used to facilitate the analysis of information contained in an image. As one of several processing techniques has been the segmentation process as simple and generally with low computation. The goal of segmentation is to isolate the structures that are being analyzed, removing what isn’t need to be considered. When applied to images of transmission electron microscopy (TEM), the segmentation of carbon nanotubes (CNT) enables the quantitative and statistical analysis of the material, providing data for the study of new materials and their properties. The segmentation can be performed using various methods that evaluate image characteristics, one is the histogram, which representes the quantitative distribution of the intensity of the pixels. This paper aims to segment CNTs on TEM images using analysis of the histogram of image regions, using various comparison histograms methods, to determine which among them has the best results given the characteristics of the problem in question. In addition to determining the method also want to determine the best size of the analyzed region fits main purpose. The methodology is divided into noise analysis, processing with different distance measures between histograms, result analysis and segmentation using defined parameters. First, the analysis of noise contained in the images will provide data for generating a statistical model thereof, which serve as the basis of comparison between the histograms. Following the application of different methods for histogram distance calculation between a region histogram and the noise model, may cite the methods of Euclidean distance, Minkowski and Bhattachyyara. It was also taken into account the size of the analyzed area, extending the analysis to determine the most suitable size for processing the histogram, and the application of the methods performed on each size tested. As a comparative measure is used segmentation analysis using the histograms distance methods with the expected result (ground truth). With work can be seen that the Manhattan distance yielded better results than other distance measures. Also it was observed that the best results within the window sizes are more frequent among 3x3 and 17x17 pixels.