Correction of Text Region Distortion for Signboard Images Recognition

Linear distortion of the text region in signboard images appears according to the photographing angle of the images and result in a reduction of the accuracy of character recognition. In this paper, we propose a method to correct the linear distortion of text regions as one of preprocessing steps. The proposed method estimates the outline of text region based on calculating the degree of the distortion by analyzing the slops of the vertical lines of characters in text regions. Then, the method corrects the distortion by a bilinear transformation to the outline of text area. The experimental results show that the similarity between artificial images and distortion-corrected images was 95.61% and the recognition rates in character and text level using the real signboard images with distortion correction is higher by 41.19% and 62.86%, respectively. In summary, the proposed method is effective to correct the linear distortion of text regions on signboard images.