In this paper we investigate influence of unsharp masking technique on OCR performance. Three different filters are used for unsharp masking: Laplacian, adaptive Laplacian and Teager filter. In order to reduce noise sensitivity we propose adaptive version of Laplacian unsharp masking: Our approach corrects output of Laplacian filter according to local variance of pixel values. Character images are taken from ICDAR 2003 dataset and enhanced with these 3 variations of unsharp masking filtering. Second step is binarization using Otsu’s criterion and processing with OCR software. Results show that approach with Teager filter shows best recognition rate. Unsharp masking based on Laplacian and adaptive Laplacian filter has weaker performance, but they still improve recognition rate in comparison with original images.

Elektrotehnika,Računarstvo