Multi-oriented text detection in scene images

Abstract

We present a new run-length based method for multi-oriented text detection in scene images. We consider one ideal Sobel edge image of the horizontal text image to compute run-lengths for multi-oriented text images. Then the method proposes a Max-Min clustering to find ideal run-lengths that represents text pixel from an array of run-lengths of ideal image. The run-lengths computed for the input multi-oriented and horizontal text images are matched with the ideal run-lengths given by the Max-Min clustering to find potential run-lengths. The boundary growing method is introduced to traverse multi-oriented text lines given by the potential run-lengths and then the method eliminates false positives to clear the background using angle-proximity features of the text blocks. The non-horizontal text image is rotated to horizontal direction based on angle of the text lines to ease the implementation. The method explore new idea based on zero-crossing to separate text lines from the touching text lines given by the boundary growing method. The proposed method is tested on our own multi-oriented scene data captured by high resolution camera and mobile camera, and the benchmark database (ICDAR 2003 competition scene images) to evaluate the performance of the proposed method. The results are compared with the existing methods to show that the proposed method outperforms the existing methods in terms of measures. © World Scientific Publishing Company.