Script pattern identification of word images using multi-directional and multi-scalable textures

As a precursor of optical character recognition (OCR) technology, script identification finds many applications like sorting and indexing of document images. Classifying these scripts, especially at different scales and orientations, is one of the interesting and vital problems in the field of document image analysis. In this paper, an algorithm is proposed for the identification of scripts using scale and rotation robust log-polar wavelet and semi decimated wavelet features. Initially, words are segmented from document images in the form of text-blobs by the Gaussian filter. Then, texture features are calculated using a combination of discrete wavelet and semi decimated discrete wavelet transforms in log-polar domain. Here, most of the rotational and scale variations are removed in log-polar domain, whereas wavelet transform is capable of extracting the information at different resolution levels. This helps in the formation of significant textures for the purpose of characterization. At last, k-nearest neighbor classifier is used for the identification of scripts. Comprehensive experiments on different databases illustrate the effectiveness of the proposed algorithm. Benchmarking analysis shows that a maximum recall rate of 98.96% is obtained, and demonstrates better performance compared to the other contemporary approaches.