English Character Recognition from Video Stream based on Bag of Visual Words (BOVW)

Abstract

Numerous digital images are available for printing the documents. The discussions are continuing for arriving to best algorithm for identifying the English letters. The suggested method has four steps. Firstly, apply wavelet transform on images of letters using Haar filter. Secondly interest points were detected using features from accelerated segment test (FAST) corner detection. Thirdly those points were descripted using Speeded up Robust Features (SURF). Fourthly, the clustering algorithm of moving k-means is employed to obtain bag of visual words (BOVW) and then build vocabulary and a histogram from visual words.The features of each visual word for video images and test image are matched using Manhattan distance measure. The suggested system was tested on three types of English letters font's databases (Time New Roman, Arial Black and Calibri) .Experimental outcomes show that the suggested method is more efficient and fast for matching and recognizing a letter than seven moment's method. The recognition time for BOVW is less than the seven moment's time and the BOVW accuracy depends on number of correct character recognition. BOVW have optimal accuracy in the process of recognition of letters.