Full text
| |
| Author(s): |
Thiago Vallin Spina
Total Authors: 1
|
| Document type: | Doctoral Thesis |
| Press: | Campinas, SP. |
| Institution: | Universidade Estadual de Campinas (UNICAMP). Instituto de Computação |
| Defense date: | 2015-09-28 |
| Examining board members: |
Alexandre Xavier Falcão;
Hélio Pedrini;
Neucimar Jerônimo Leite;
Silvio Jamil Ferzoli Guimarães;
Luís Gustavo Nonato
|
| Advisor: | Alexandre Xavier Falcão |
| Abstract | |
With the rise of social media, the behavior of regular users has changed from merely consuming to actively producing multimedia data content that is shared on-line. In this scenario, several applications have been developed for photo and video editing. Interactive image and video object segmentation are often needed for those applications, demanding for effective and efficient methods that help the user to extract the objects of interest from the background accurately, while requiring minimum user effort and time. Segmentation may be divided into object recognition and delineation. Recognition involves approximately locating the object and verifying the segmentation result, being a simpler task for humans. Delineation involves defining the object's spatial extent in images and video frames, which can be done more precisely by computers. Interactive methods seek a synergy between the user and the machine, by propagating the labels from the user's annotations (i.e., scribbles, regions of interest, contour initializations in images and video frames) to the unlabeled data. In image segmentation, a natural way of exploiting the connection between the object information provided by the user (scribbles) and the pixels is to consider the image explicitly or implicitly as a weighted graph. Several graph-based frameworks may be used for delineation, but a challenge is to estimate arc weights that make segmentation trivial. The first main contribution of this PhD thesis is a methodology for enhancing the differences between foreground and background to aid in arc-weight assignment, which considers the user's input for supervised learning in a transparent way to the human operator. Our second contribution intends to make user interaction more effective in the presence of imperfect arc-weight estimation. It involves the development of hybrid interactive image segmentation techniques that combine region-based methods with boundary-tracking approaches to explore the advantages of both. The former typically handle complex silhouettes more easily, while the latter allows the user to select accurate boundary segments to compose the object's contour. Although the aforementioned interactive image segmentation tools could be used frame-by-frame, it is more effective to develop approaches that automatically propagate the object information from an input frame to the rest of the video. Our third contribution involves embedding the user's knowledge about the object into a fuzzy shape model for object recognition. This model aims to minimize the need of human intervention by automatically correcting segmentation, as propagated to new frames. The user may still refine the result with our image segmentation tools when necessary (AU) | |
| FAPESP's process: | 11/01434-9 - Interactive segmentation of objects in digital video using graphs and fuzzy models of content knowledge |
| Grantee: | Thiago Vallin Spina |
| Support Opportunities: | Scholarships in Brazil - Doctorate (Direct) |