Text Detection and Recognition from Scene Images using MSER and CNN

Abstract

Detection and recognition of text from natural images is very important for extracting information from images but is an extensively challenging task. This paper proposes an approach for detection of text area from natural scene images using Maximally Stable Extremal Regions (MSER) and recognizing the text using a self-trained Neural Network. Some preprocessing is applied to the image then MSER and canny edge is used to locate the smaller areas that may more likely contain text. The text is individually isolated as single characters by simple algorithms on the binary image and then passed through the recognition model specially designed for hazy and unaligned characters.