Text string detection for loosely constructed characters with arbitrary orientations

Abstract

Text in a scene provides vital information of its contents. With the increasing popularity of vision systems, detecting general text in images becomes a critical yet challenging task. Most existing methods have focused on extracting neatly arranged text string for compactly constructed characters. Motivated by the need to consider the widely varying forms of scene text, we propose a stroke-based text detection method which detects arbitrary orientations text strings with loosely constructed characters in images. Our approach employs result of stroke width transform (SWT) as basic stroke candidates. These candidates are then merged using adaptive structuring elements to generate compactly constructed characters. Individual characters are chained using k-nearest neighbors algorithm to identify arbitrary orientations text strings, which are subsequently separated into words if necessary. To better evaluate our system and compare it with other competing algorithms, we generate a new dataset, which includes various characters and text strings in diverse scenes. Experiments on ICDAR datasets and the proposed dataset demonstrate that our approach compares favorably with the state-of-the-art algorithms when handling arbitrary orientations text strings and achieves significantly enhanced performance on loosely constructed characters in scenes.