A Novel spatial relational features generation using encoded geometry for character recognition

This paper proposes two different feature extraction techniques for small-size datasets in the context of character recognition. The aim is to achieve high accuracy even with limited training data. The proposed feature extraction techniques are contour-based feature vector (CFV) and tracing line-contour intersection count (TCIC). The proposed algorithms were trained using a small data set containing English alphabets and numerals. The feature extraction techniques were used to extract relevant features from the characters in the training set, and the resulting feature vectors are used to classify the input character by the Levenshtein edit distance classifier. The trained model was then tested on the test set to evaluate its accuracy. Experimental results shows that the proposed approach achieved encouraging results on publicly available MediaLab data set with an overall accuracy of 97%. The best performance was obtained using a combination of proposed algorithm and crossing count feature vector. The results demonstrate that even with a small datasets, high accuracy can be achieved using appropriate feature extraction techniques and a well-designed classifier. In addition, the proposed methods were able to accurately classify broken and distorted characters, which are often challenging to recognize using conventional approaches. These findings can be useful for developing offline character recognition systems for embedded devices and mobile devices with very limited computing power.