A robust algorithm for text extraction from images

Abstract

A robust algorithm that detects text from natural scene images and extracts them regardless of the orientation is proposed. All existing methods are designed to operate under a certain constraint, like detecting text only in one direction. Maximally Stable Extremal Regions (MSER) detector is chosen to extract binary regions since it has proven to be robust to lighting conditions. An enhancement technique for MSER images is designed to obtain clear letter boundaries. Images are then fed into a Stroke Width Detector and several heuristics are applied to remove non-text pixels. Afterwards, detected text regions are fed into an Optical Character Recognition module and then filtered according to their confidence measure. The recognition of characters is not part of the algorithm and the results are only about the detection of text. Our algorithm proved to be effective on blurred images and noisy images as well, based on both subjective and objective evaluations.