Text localization and extraction from background with texture and noise in digital images using adaptive thresholding and convolutional neural network

Abstract

For the past few years, research topics on finding position of text have received more and more attention from researchers because there are still lots of problems that are needed to be solven. We propose a novel method to find the position of text in an image. The first step of the proposed method is to adjust an image by converting a color image to a gray scale image and then use an average filter to improve an image. An average filter is used to make the background smooth and reduce noise. After that an adaptive thresholding is used to convert a gray scale image to a binary image. In the second step, the 8-neighbor connected component is used to generate bounding boxes to find the position of text candidates and the size of a bounding box is used to classify a bounding box as a text candidate or a non-text candidate. In the third step, the width-to-height ratio and Ostu’s thresholding are used to divide text candidate into character candidates. Then convolutional neural network is used to categorize a character candidate as character or non-character. The proposed method was evaluated on a data set from ICDAR 2013 which contains 229 training image data and 233 test image data. The classification revealed that the proposed method yields 71.87% of recall and 36.59% of precision. This research can find bounding box of character effectively but it can not delete bounding box of non-character well.