Filtros espaciais: estudo comparativo e aplicação em classificação e segmentação de imagens.

Abstract

This paper addresses the problem of automatic discrimination of cardiac tissues on ultrasound textural, and image enhancement techniques can be used to improve the results of such discrimination. Initial results, obtained through the use of of textural analysis together with the discriminant analysis step by step, indicated a rate of up to 95% success in discrimination of fabrics. The results of this study refer to called the learning phase of an automatic classification system, where it is sought to build a knowledge base on the textural parameters that best lend themselves to discrimination. With the objective of improving these results, image enhancements, implemented by spatial filters, were applied as pre-processing algorithms, before execution of the textural analysis. The filters to be used were selected using a systematic study of techniques already reported in the literature and a performance analysis involving those selected techniques. In the performance evaluation of the algorithms are considered noise removal capabilities with edge preservation and impulsive noise, ramp type sharpening, preservation of thin details and fine characteristics, as well as its computational efficiency. As a result of the study of spatial filters, a new filtering technique was developed for smoothing by elective neighborhoods. Compared to other algorithms analyzed this technique proved to be an efficient solution to the problem of noise removal with edge preservation. The application of spatial filters as preprocessing algorithms in the task of discriminating between diseased and healthy tissues in cardiac ultrasonography, resulted in an increase 100% success rate in discrimination. The results of this experiments are provided in the form of statistical parameters, contingency tables and two-dimensional diagrams that show the separation of classes. In addition, we investigated the use of preprocessing techniques in automatic image segmentation tasks, aiming at the problem of determining the contours of objects. Preprocessing of the images resulted in a reduction number of parameters (edges and regions) detected and the number of performed.