Full text
| |
| Author(s): |
Maíra Saboia da Silva
Total Authors: 1
|
| Document type: | Master's Dissertation |
| Press: | Campinas, SP. |
| Institution: | Universidade Estadual de Campinas (UNICAMP). Instituto de Computação |
| Defense date: | 2011-07-25 |
| Examining board members: |
Alexandre Xavier Falcão;
Silvio Jamil Ferzoli Guimarães;
Hélio Pedrini
|
| Advisor: | Alexandre Xavier Falcão |
| Abstract | |
This work presents a new methodology for automatic extraction of desired objects in natural images. Objects of interest are defined as the largest components that differ from their surrounding pixels in a given image. These objects do not need to be centered, but they should contain a minimum number of pixels in the region assumed as background (e.g., an image border of certain thickness). This methodology is based on background segmentation and it can be summarized in two steps. First, a fuzzy model is created by a clustering method based on probability density function of the background colors. This model is a membership map, wherein object pixels are brighter than background pixels. For clustering, the following techniques were investigated: mean-shift, image foresting transform, Gaussian mixture model and expectation maximization. We then propose and compare three approaches to create a membership map; a first method entirely based on the image foresting transform, a second approach based on Gaussian mixture model and a third tecnique using expectation maximization. The clustering based on image foresting transform was adopted as the initial estimate for the clusters in the case of the two last methods. In a second step, the membership map is used to enable the selection of object and background seed pixels. These pixels create a binary clustering of the color pixels that separates background and object(s). The experiments involved a heterogeneous dataset with 50 natural images. The approach entirely based on the image foresting transform provided the best result. In order to justify the use of a binary clustering of color pixels instead of optimum thresholding on the membership map, we demonstrated that the binary clustering can provide a better result than Otsu's approach. It was also proposed a hybrid approach to binarize the membership map, which combines Otsu's thresholding and image foresting transform. In this case, Otsu's thresholding reduces the number of parameters in regard to the first approach (AU) | |
| FAPESP's process: | 10/04936-2 - PIXELS CLUSTERING BY OPTIMUM PATH FOREST FOR IMAGE SEGMENTATION |
| Grantee: | Maíra Saboia da Silva |
| Support Opportunities: | Scholarships in Brazil - Master |