A Hybrid Approach to Detect Text and to Reduce the False Positive Results in a Scenery Image

Abstract

Detection of the text from the scenery images containing text is a challenging task that has received a lot of attention recently. In scenery images there are two key components 1) finding text from images, 2) Recognition of character. Many researchers have published their work on both components. Finding the text in the images is the primary part because the overall accuracy of the model depends on the output of this phase. In the present study, a method has been proposed that consists of two phases 1) Text detection 2) Text verifier. Text detection is done through the well-known algorithm for text detection-MSER (Maximally Stable Extremal Regions) feature detector. Then different filters like elimination of non-text region based on simple geometric properties, and elimination of non-text region based on stroke width variationon the output of MSER feature detector are applied to filter out the components that possibly cannot be the text. In second phase, machine learning approach (ANN-classifier which acts as text verifier) is used to classify the text and non-text on the final output of phase 1. It is found that proposed algorithm almost eliminate all the false positive results on the final output of the MSER feature algorithm.