Análise de imagens por meio da matriz de interdependência e da transformação estrutural multiescala

Abstract

Image analysis is a fundamental task in computer vision. It influences the development of algorithms for digital image processing and approaches for evaluating the results produced by these algorithms. This thesis introduces a methodology for the structural analysis of images based on the combined use of a multiscale structural transformation and extraction of structural features through spatial interdependence matrix. The multiscale structural transformation is an algorithm based on mathematical morphology framework that maps the gray levels of the input image into a space in which these gray levels are grouped into different scales of structures that form objects. The transformation can be applied in enhancement of gray level images and decomposition of binary images into elementary structures. The spatial interdependence matrix is an algorithm based on cooccurrence statistics that produces a global representation of the structural coincidences of two images. This matrix provides four attributes, namely, correlation, inverse difference moment, chi-square coefficient and entropy, which can be used as global descriptors of the image structures. The proposed methodology is validated with the results obtained for different applications: the detection of atmospheric corrosion of metal surfaces in photographs, the detection of lung disease in computerized tomography images, the referenced evaluation of image quality, the segmentation of retinal vessels in retinography and the quality assessment of retinal vessels segmentation algorithms.