Nombre: BRUNO LÉGORA SOUZA DA SILVA
Fecha de publicación: 08/12/2016
Supervisor:
Nombre |
Papel |
|---|---|
| PATRICK MARQUES CIARELLI | Advisor * |
Junta de examinadores:
| Nombre |
Papel |
|---|---|
| KELLY ASSIS DE SOUZA GAZOLLI | External Examiner * |
| KLAUS FABIAN COCO | External Examiner * |
| PATRICK MARQUES CIARELLI | Advisor * |
Sumario: Text localization in real world images is a problem which has received significant attention in the present days, since it can be used in various applications, such as geolocation, indexing of images, identification of products through their labels and assistance to the visually impaired, improving their quality of life. Currently, there are several researches in this area, WHERE a competition held in the International Conference of Document Analysis and Recognition follows the progress of these researches, and this highlights the relevance of this area. This work presents some techniques of text segmentation in real world images, WHERE text occurs in several forms, orientations and sizes. Such techniques are based on connected component analysis, using edge detection, techniques such as Maximally Stable Extremal Regions, the FASText technique that finds keypoints based on the FAST detector, techniques based on sliding windows or the combination of these two approaches. An experimental comparison of three different text segmentation techniques is done. These have a reduced computational cost, so that they do not have a major impact on the computational cost of a system that uses them. For this comparison, three databases are used: the two divisions of the ICDAR 2013 competition database and the KAIST Scene Text database, analyzing metrics such as extracted text area, processing time, number of characters without segmentation and the imprecision of the detector, defined as the ratio of the number of targets to the number of characters in the database. The results obtained in the databases have shown that all the analyzed techniques obtained a good segmentation ratio in the databases, when the images presented well defined characters, without problems of illumination, occlusion or very small characters. In the images that present the latter, the analyzed techniques obtained impared performances.
