Analysis of the Proposed CNN Model for the Recognition of Gurmukhi Handwritten City Names of Punjab

A holistic approach employing different learning rates is proposed for the recognition of handwritten district names of Punjab state which are written in Gurmukhi Script. Learning rate (LR) is a hyperparameter that specifies how much the model can change when the model parameters are changed in terms of error. For the purpose of recognition, a convolutional neural network (CNN) using deep learning is proposed. Initially, the dataset of 4000 of images is prepared for ten different city names of Punjab state, and later, a CNN is employed. The proposed CNN architecture having 12 layers is developed and employed using five different learning rates: 0.00001, 0.0001, 0.001, 0.01 and 0.1, and their performance is analysed for the purpose of text recognition. Best average validation accuracy achieved for the proposed CNN architecture is 93%, and maximum achieved validation accuracy is 99% by employing LR of 0.0001 for the prepared dataset.