Self- attention based optimized deep convolutional robust character and numeral recognition from Gujarati language

In this research work, character and numeral recognition from the Gujarati language is proposed with the aid of an optimized deep convolutional network. In the pre-processing stage, the binarization, filtering, and morphological operations are performed on the characters/numerals. In the second stage, the diagonal, centroid, peak extent, and zoning features are extracted. In the feature fusion stage, features are integrated by the technique known as canonical correlation analysis (CCA). Finally, in the recognition stage, the self-attention based honey badger convolutional (SA-HBCNN) is proposed to obtain accurate recognition results for the character and numerals classification from the handwritten Gujarati language. The ablation study is conducted for the proposed SA-HBCNN model with and without pre-processing stages, and the performance measures of accuracy, precision, recall, and f-measure are also examined. The achieved performance for character recognition is 99.84%, 99.23%, 99.22%, and 99.23%. The outcomes for numeral recognition are 99.81%, 99.05%, 99.08%, and 99.07%.