Natural images often contain useful information

In this paper, a new method of binarizing degraded characters from scene images is presented. The proposed method helps in obtaining a binarized result even with multicolored characters subjected to heavy degradation. Color information is used in the proposed binarization method to help to separate the character from the image background and to obtain a clean representation of the final result. Images from the ICDAR 2003 robust character recognition database are used to compare the effectiveness and accuracy of the proposed algorithm with other methods.

Comparison of image segmentation techniques based on the number of dominant colors and clusters is presented. Tensor Voting, Expectation Maximization algorithm, K-Means Algorithm and Mean Shift Algorithm are considered. The image segmentation results are analyzed with constant and varying number of clusters for all algorithms. Finally the performance of all algorithms under Gaussian noise is also evaluated. Performance results suggest that Tensor Voting based segmentation is more robust to noise compared to other techniques.