Modified Stroke Width Transform for Thai Text Detection

Abstract

Stroke width transform (SWT) is widely used for detecting the candidate text in a natural scene image before forwarding it to character recognition to convert to editable text. It is an important part of an OCR application. However, the main problem affecting its performance is inconstant stroke width. This problem originates from inaccurate stroke width map. Due to this inaccurate map, the candidate text may get rejected in the SWT filtering step even though it is so. This leads to lower SWT performance. In order to solve this problem, we propose a modified stroke width transform (MSWT). This method divides a stroke width map from the candidate object into several sub-regions and computes the variance and mean of each sub-region instead of those of the whole stroke width map. This can reduce the destructive effect from inaccurately computed stroke width. Based on 100 images of natural scene with embedded Thai text, an experiment was conducted to comparatively evaluate the performance of MSWT against SWT in Thai text detection. The experimental results show that the recall rate of the proposed method was higher than that of the conventional SWT method.