view in publisher's site
- لیست مقالات
Genetic Algorithm Based Global and Local Feature Selection Approach for Handwritten Numeral Recognition
Although various methods have been proposed by the researchers over the years to carry out isolated handwritten numeral recognition, this is still considered as a challenging research problem. The primary challenge occurs because of sizable differences in writing styles of the digit patterns. Literature reveals that several feature extraction and classification methods have been researched upon to optimize the said recognition system but there is still room for improvement. Here, we introduce a two-stage handwritten numeral recognition system for three most popular scripts used in Indian subcontinent—Devanagari, Bangla and Roman. Initially, the global features that are estimated based on Histogram of Oriented Gradients (HOG) feature descriptor help in the formation of inter-numeral groups having nearly similar structure. Then, the optimal subset of features is selected using Genetic Algorithm (GA) on the combination of HOG and local distance features. These optimal features, thus produced, are employed for the classification of handwritten digits within the intra-numeral groups. Finally, the Multi-layer Perceptron (MLP) classifier is used to recognize the numerals. The strength of the present approach is efficient feature selection and the comprehensive classification scheme due to which notable recognition accuracies have been attained which are found to be better than some previous approaches.