Localization and Recognition of Text with Perspective Distortion in Natural Scenes

Recognizing text in natural scene images refers to the problem of identifying words that present on it. Scene text recognition is very difficult due to some reasons such as, images contain very little amount of linguistic context, interpreting versions of letters and digits are required for scene text recognition and also scene text can appear in any orientation. Most of the existing works are focused on the recognition of texts which are frontal parallel to the image plane. We formulate a novel method which is used to recognize text in natural scene images which are perspectively distorted. Perspective distortion is avoided using Hough transform. Each character are recognized from cropped word image. Connected component analysis is used to detect the components that present on the cropped word image. Non text components are filtered using SVM classification. After that text components are recognized by Optical character recognition. We introduce a new dataset called Scene Text-Perspective, which contains scene images of the name boards placed in the road sides which are perspectively distorted. Experimental results on the proposed dataset shows that our method is simple and outperforms the existing methods.