Text Image Classifier Using Image-Wise Annotation

Abstract

A text image classifier that requires only image-wise annotation is proposed. Although text detection methods using classifiers have been investigated, they require character-wise annotation by human operators, which is the most time-consuming phase when constructing a text detection system. The proposed classifier uses image-wise annotation whether the image contains text or not, which requires much less effort by an operator than that of character-wise annotation. From this annotation, the classifier estimates likelihood of detecting text-character candidates in an image as well as the threshold value for the system to determine if the image contains text based on prior probabilities. Experiments using real images showed the effectiveness of the proposed text image classifier.