Character shape restoration system through medial axis points in video
Shape restoration for characters in video is challenging because natural scene characters usually suffer from low resolution, complex background and perspective distortion. In this paper, we propose histogram gradient division and reverse gradient orientation in a new way to select Text Pixel Candid...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Published: |
Elsevier
2015
|
| Subjects: | |
| Online Access: | http://eprints.um.edu.my/19541/ http://dx.doi.org/10.1016/j.neucom.2015.02.044 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Shape restoration for characters in video is challenging because natural scene characters usually suffer from low resolution, complex background and perspective distortion. In this paper, we propose histogram gradient division and reverse gradient orientation in a new way to select Text Pixel Candidates (TPC) for a given input character. We apply a ring radius transform on TPC in different directions, namely, horizontal, vertical, principal and secondary diagonals in a TPC image to obtain respective radius maps, where each pixel is assigned a value that is the radius to the nearest TPC. This helps in finding Medial Axis Points (MAP) by searching for the maximum radius values from their neighborhoods in a radius image. The union of all the medial axis points obtained from the respective directions at each location is considered as Candidate Medial Axis Points (CMAP) of the character. Then color difference and k-means clustering are proposed to eliminate false CMAP, which outputs Potential Medial Axis Points (PMAP). We finally propose a novel way to restore the shape of the character from the PMAP. The method is tested on a video dataset and the benchmark ICDAR 2013 dataset to show its effectiveness for complex background and low resolution. Experimental results show that the proposed method is superior to the existing methods in terms of shape restoration error and recognition rate. |
|---|