Optical character recognition in images of natural scenes

Rok Petek (2012) Optical character recognition in images of natural scenes. EngD thesis.

Preview

PDF Download (2586Kb)

Abstract

The diploma thesis presents a description and implementation of some modern techniques and methods for optical character recognition in images of natural scenes. When choosing methods, we focused on speed and accuracy. As a basis we have chosen the method of directional segment features with nonlinear mesh since it was developed for the mobile platform and thus meets the criteria of speed. Also, the method comparable to other methods reaches very good results. The proposed method was further upgraded with some other popular features extraction methods and classifiers. Optical recognition of text in images of natural scenes is very problematic, because in them the text appears in a variety of sizes, colors, fonts and orientations. Also pictures of natural scenes typically have lower quality and contain complex background, which greatly complicates the process of recognition. Similar to the classic optical character recognition the systems for optical character recognition in the images of natural scenes typically consist of four steps: preprocessing, segmentation, feature extraction and classification. Preprocessing phase is designed to improve image quality, in the segmentation stage only the pixels that belong to each character are chosen in the picture. Both steps are due to the aforementioned problems of natural scenes extremely important. During the feature extraction phase the characteristics of a segmented character are calculated, which are user for further classification of the character corresponding class. All implemented methods were tested on image databases ICDAR, CVL OCR DB, and a hybrid collection that we have generated from the two mentioned databases. Improved method presented in the thesis has achieved good results and is, in conjunction with the relevant text detection in images of natural scenes, suitable for migration and the use on the mobile platform.

Actions (login required)