Métodos auto-organizáveis para segmentação de imagens

Abstract

Image segmentation is one of rnost investigated problems in computer vision. Its complexity can vary according to the kind of application. In general, the goal is to divide an image into regions with similar properties. In tliis work, self-organizing methods for unsupervised classification and clustering are applied in image segmentation tasks. The first self-organizing model is the Fuzzy ART neural network and the other one is the ICA Mixture Model (ICAMM), which uses ICA method to desenhe data in each class. Beside the performance evaluation regarding the considered methods, some improvements on the segmentation results obtained by these techniques were proposed by incorporating some image preprocessing methods. Sueli methods were able to handle some questions regarding to presence of noise, image smoothing and edge enhancement, in a way that makes an image more suited to be processed by an image segmentation technique, which can become more efficient. Aiming this, a preprocessing methodology was proposed in this work that combines Sparse Code Shrinkage method for image denoise to the Sobel Edge Detector, which is applied to recover edges that were blurred by an excessive smoothing. Another original contribution of this work refers to the development of EICAMM, which was built by proposing some modifications on ICAMM, considering some limitations on the original method and analysis on liow it should be modified to become more efficient. Finally, unifying the two main contributions of this thesis, the EICAMM method was applied for segmenting some images in these original and preprocessed versions, obtained by the proposed preprocessing methodology. Such systern lias showed satisfactory image segmentation results.