Propagação de etiquetas para segmentação de imagens

Abstract

Image segmentation is an important task in image processing, usually employed in more complex computer vision tasks. In graph clustering-based segmentation approaches, the image is modeled by a graph, in which vertices are generally represented by pixels and edges by weights that denote similarity between pixels. The problems associated with graph-based approaches usually concern the computational cost and the high cardinality of the graphs, which translates into the large number of vertices and edges necessary to generate an adequate representation of the image. Among segmentation approaches with graphs, those based on detection of communities in complex networks, such as Label Propagation, in particular because they have a lower computational cost, stand out. However, such methods when applied directly to images, do not generate accurate results, in addition to being non-deterministic, which is an undesirable quality in image segmentation. On the other hand, superpixels techniques, which combine several pixels, are important not only in reducing the cardinality of the graphs, but also in providing greater descriptive power compared to a single pixel. This doctoral thesis presents a new family of segmentation methods for images of large natural scenes images based on the Label Propagation and superpixels method, with deterministic behavior and which uses specific information from the image domain. Algorithms were developed for both automatic segmentation (SGLP-Simple Graph Label Propagation and MGLP-Multi-level Label Propagation), and for interactive segmentation (IGLP-Interactive Graph Label Propagation), which demand user assistance. Quantitative results show a PRI precision of 0.83 and error percentage of Er 6.13%, for the automatic and interactive version, respectively. Results were also obtained in the processing time of 0.0048 s and 0.29 s, for automatic and interactive segmentation. These results were corroborated in several experiments on standard data sets. When compared with related methods, the results of the methods are superior both in mean precision and time for automatic segmentation. As for the interactive segmentation method (IGLP), segmentation mean precision was slightly outperformed by state of the art methods, but run in shorter times.