Scene Text Extraction by Combining Edge Based Stroke Segmentation and Morphological Filtering

Extraction of text embedded in scene images has many applications, such as license plate recognition, content based retrieval, text based indexing etc. As the appearances of the text in images are unpredictable, separating text pixels from the background is a challenging one. A new frame work is introduced here which uses edge enhanced MSER along with connected components for text extraction. Complementary features of MSER and sobel edge methods are combined to obtain the better results. Feature vector obtained from the stroke width transform are used to recover the text connected components. Further morphological operations along with heuristic filters are used to separate the text pixels from the background. Validation of the method proposed is tested on ICDAR datasets and also on own dataset. The experimental outcome indicates that the suggested algorithm extracts the text characters efficiently even when there is a variation in font, size, color, under different conditions such as low resolution, complex background, varying illuminations etc.