Script Identification from Camera Based Tri-Lingual Document

In this paper, an algorithm is proposed for Trilingual Script Identification System in block wise for camera captured images. The Local Binary Pattern (LBP) features are used for Kannada, Hindi and English images for testing the performance of a proposed algorithm, a dataset of 6000 neat block images are considered. For each script a total of 2000 images are used for the proposed method. The segmentation technique is used to segment the document image in blocks. Block of sizes 128x128, 256x256, 512x512 and 1024x1024 for Kannada, Hindi and English have been considered. The LBP features are extracted in 8 neighbors, there by generating 59 features and submitted to KNN and SVM classifiers to classify the underlying image. The identification accuracy for KNN and SVM classifiers are respectively 96.60% and 98.00% for block size 128x128, 98.71% and 98.07% for block size 256x256, 99.70% and 98.00% for block size 512x512 and further 94.90% and 99.01% for block size 1024x1024 respectively. The optimal accuracy is 99.01% for SVM classifier for block size 1024x1024. The proposed method is independent of thinning.