Segmentação não supervisionada de texturas baseada no algoritmo PPM
The image segmentation problem is present in various tasks such as remote sensing, object detection in robotics, industrial automation, content based image retrieval, security, and others related to medicine. When there is a set of pre-classified data, segmentation is called supervised. In the case...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Estado: | Versión publicada |
| Fecha de publicación: | 2010 |
| País: | Brasil |
| Institución: | Universidade Federal da Paraíba (UFPB) |
| Repositorio: | Biblioteca Digital de Teses e Dissertações da UFPB |
| Idioma: | portugués |
| OAI Identifier: | oai:repositorio.ufpb.br:tede/6147 |
| Acceso en línea: | https://repositorio.ufpb.br/jspui/handle/tede/6147 |
| Access Level: | acceso abierto |
| Palabra clave: | Segmentação não supervisionada Análise de texturas PPM Algoritmo aglomerativo guloso Agrupamento Unsupervised segmentation Texture analysis Greedy agglomerative algorithm Clustering CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO |
| Sumario: | The image segmentation problem is present in various tasks such as remote sensing, object detection in robotics, industrial automation, content based image retrieval, security, and others related to medicine. When there is a set of pre-classified data, segmentation is called supervised. In the case of unsupervised segmentation, the classes are extracted directly from the data. Among the image properties, the texture is among those that provide the best results in the segmentation process. This work proposes a new unsupervised texture segmentation method that uses as the similarity measure between regions the bit rate obtained from compression using models, produced by the Prediction by Partial Matching (PPM) algorithm, extracted from them. To segment an image, it is split in rectangular adjacent regions and each of them is assigned to a different cluster. Then a greedy agglomerative clustering algorithm, in which the two closest clusters are grouped at every step, is applied until the number of remaining clusters is equal to the number of classes (supplied by the user). In order to improve the localization of the region boundaries, the image is then split in shorter regions, that are assigned to the cluster whose PPM model results in lower bit rate. To evaluate the proposed method, three image set were used: Trygve Randen, Timo Ojala and one created by the author of this work. By adjusting the method parameters for each image, the hit rate obtained was around 97% in most cases and 100% in several of them. The proposed method, whose main drawback is the complexity order, is robust to regions with different geometric shapes, grouping correctly even those that are disconnected. |
|---|