Text Localization and Extraction from Background with Texture and Noise in Digital Images Using Adaptive Thresholding and Convolutional Neural Network

Recently, the task of text localization attracts many attentions. In this paper, we propose a method to localize individual characters of text from background with texture and noise in digital images using adaptive thresholding, width-to-height ratio, and convolutional neural network. The proposed method consists of three main steps: preprocessing, text candidate localization and classification, and character candidate localization and classification. In preprocessing step, images are enhanced by average filter in order to reduce noise and make the texture background harmonious. Then text candidates in an image are localized and classified using adaptive thresholding and width-to-height ratio of bounding boxes. Finally, character candidates of each text are localized and a convolutional neural network is used to classify character candidates as either character or non-character. The experiments were conducted on a dataset obtained from ICDAR2013 containing the training set of 229 images and the testing set of 233 images. From the experimental results, the proposed method can correctly localize text with the accuracy of 71.87%.