Skip to Main Content
In this paper, a new approach for text detection and localization is proposed. For this purpose, we first localize text location and then determine characters' pixels. The proposed text detection approach is a two-stage algorithm that in first stage, we apply low pass filter on image in FFT domain to remove noisy element and then we apply Laplacian operator to the resultant image to highlight high contrast areas in the image. Then the product of corner dilated points and Laplacian enhanced mage is calculated and text blocks are extracted using image vertical and horizontal projection. In the second stage of the algorithm, the extracted text blocks are verified using an SVM classifier. Text textures such as text angles and variance, momentum, entropy in co-occurrence matrix of text block are used for SVM training. We assumed that the characters of each text block have the same color. Therefore, we first estimate background color using image pixels in borders of detected text areas. Then the text color is estimated using the color clusters of pixels in text block and background color. We use color segmentation to extract character pixels. Experimental results show the promise of the proposed algorithm.
Date of Conference: 6-8 Nov. 2012