Automatic localization and recognition of perspectively distorted text in natural scene images
Annmaria Cherian, Sanju Sebastian · 2016
The automated understanding of textual information from the image is the main goal of Scene Text Recognition (STR). STR is very difficult due to several reasons such as viewing angle and lighting, which are not carefully controlled and very little amount of linguistic context contained in scene images due to which other objects present in the image can interfere the recognition process. Most of the existing work is focused on the recognition of texts which are frontal parallel and horizontal to the image plane. We formulate a novel method to recognize text in natural scene images which are perspectively distorted. Our method uses the Hough Transform to correct the orientation of scene images and uses efficient character detection and localization technique. SVM classifier is used to filter the non-text components from the detected components. After filtering, character recognition is used to recognize the text accurately. Here a new dataset called Scene Text-Perspective is introduced, which contains scene images of the name-boards placed on the road sides which are perspectively distorted. Experimental results on the proposed dataset shows that our method is simple and outperforms the existing methods.