Recognition of Text-based CAPTCHA with Merged Characters

Abstract

This paper intends to introduce an approach based on generalized Hough transform with the aim to recognize the CAPTCHA with merged characters. Firstly, pre-processing and skeletonization are applied to the CAPTCHA. In the meantime, the local features of the edge points in template image are extracted and saved in the reference table. During the GHT voting phase, each object pixel of the CAPTCHA skeletonization image is compared with the reference table to accumulate the voting array. Therefore, the coordinates of the potential reference points are located by finding the maxima of the array. Finally, the interference of neighboring characters is analyzed to heighten the recognition rate. Due to the shift and rotational invariance of the local features, this method can effectively solve the recognition of slightly merged CAPTCHA with local shape variation and rotational characters.