Word-wise Script Identification in Document Images based on Steerable Gaussian Filtering Technique

Abstract: In this paper, a study on word wise script identification based on Steerable Gaussian filter for printed document images is carried out. The system is developed and tested for 3000 document image data set representing English, Hindi, Kannada, Tamil and Urdu script word images of 600 each. The system developed includes a feature extractor which is based on Steerable Gaussian filter technique and for classification K-nearest neighbor classifier and linear discriminate classification techniques are used. The feature extractor consists of application of steerable Gaussian filter at different orientations 0, 45, 90, 135, 30, 65, 155 and the associated standard deviation of the local orientation is used as the feature set thus contributing only seven features. The two classifications techniques were used for analysis of the new word-wise segmented documents. Classification accuracy averaged 97% across the five scripts .The method shows robustness with respect to noise, the presence of headlines, font sizes and styles.

Keywords: Document image processing, steerable filter, script, Identification, OCR