Word Dataset (Sword6k)

A Roman word dataset of scene word text images named Scene Word Dataset (SWord6k) is developed for character segmentation, Character recognition, text detection, and script identification. All images of SWord6k datasets are captured from an outdoor environment. These images are collected from banners, advertisements, shop names, and posters from different sources, like shopping malls, book fairs, and puja pandals. The SWord6k dataset contains 6,661 scene word images in "png" file format. Three types of ground truth annotations are composed for the SWord6k dataset, viz (i) component level i.e. to check whether a component is text or not, (ii) Script level i.e. to identify the script, (iii) recognition level i.e., character/word recognition of text. Each image's ground truth level annotations are stored in XML(extensible markup language) file format.