Degraded character recognition from old Kannada documents

This paper addresses preparation of a dataset Kannada characters which are degraded and robust recognition such characters. The proposed algorithm extracts the histogram oriented gradients (HOG) features block sizes 4x4 8x8 followed by principal component analysis (PCA) feature reduction. Various classifiers experimented with fine K-nearest neighbor classifier performs best. performance model is evaluated using 5-fold cross validation method receiver operating characteristic curve. devised size 10440 having 156 classes (distinct characters). These from 75 pages not well preserved old books. A comparison other like Haar wavelet Geometrical suggests that superior. It observed PCA reduced resulted in best accuracy acceptance rate 98.6% 97.9% for respectively. experimental results show HOG extraction has high system even extensively