23 April 2020
16:00 Master's Defense Fully distance
Theme
Generation of Superpixels by the Iterative Generating Forest using Object Information
Student
Felipe de Castro Belém
Advisor / Teacher
Alexandre Xavier Falcao
Brief summary
Superpixel image segmentation methods aim to partition the image into related regions (\ ie superpixels) so that the objects of interest are represented by the union of their superpixels. This result is extremely important for countless applications, increasing computational performance and allowing to explore intermediate level information about the objects involved in the image analysis. Depending on the algorithm, the performance of the segmentation of superpixels can be proportional to the number of regions generated. However, the lack of information about the objects of interest means that effective results are almost always related to unnecessary over-segmentation, negatively affecting the aforementioned purposes. In view of the development of efficient methods for segmentation of superpixels, whose objects are represented effectively with few superpixels, this work incorporates the object information in the framework of the Iterative Generating Forest (ISF, from the English \ emph {Iterative Spanning Forest}) . The resulting framework, called Iterative Generating Forest based on object information (OISF, from the English \ emph {Object-Based ISF}) consists of three independent steps, similar to the ISF: (i) initial sampling of seed pixels; (ii) delineation of superpixels from seeds using the Image-Forest Transform ~ (IFT, from English \ emph {Image Foresting Transform}) algorithm for a given connectivity function (\ ie path cost in an image graph); and (iii) recomputing of the seeds, followed by iterative execution of steps (ii) and (iii) to improve the location of the seeds and, consequently, the design of the superpixels. The object information comes from a projection map, which is previously generated and used to incorporate object information in the three stages of the OISF. The results include greater efficiency in the design with a significantly smaller number of superpixels and flexibility in adapting the framework for different applications. These results are demonstrated in comparison with several other state-of-the-art methods, including ISF-based methods, using two natural image bases and one medical image base.
Examination Board
Headlines:
| Alexandre Xavier Falcão | IC / UNICAMP |
| Paulo André Vechiatto de Miranda | IME / USP |
| Hélio Pedrini | IC / UNICAMP |
Substitutes:
| Sandra Eliza Fontes de Avila | IC / UNICAMP |
| Thiago Vallin Spina | LNLS - National Synchrotron Light Laboratory |