Towards Optimal Convolutional Neural Network Parameters for Bengali Handwritten Numerals Recognition

This work attempts to find the most optimal setting for a convolutional neural network (CNN) for Bengali digit dataset classification. Recognition of handwritten Bengali numerals has recently gained much interest among researchers due to the significant performance gain found in the recognition of English numerals using neural network based architecture. In this work, a new dataset of 70,000 samples were created first by taking handwriting of 1750 persons where 982 persons were male and the rests were female. These individual image samples are then converted to grayscale, normalized, inverted and pickled to complete the data preprocessing step. Later this dataset was recognized using several convolutional neural network settings where the most optimal setting being found to be two convolution layer with Tanh activation, one hidden layer with Tanh activation and one output layer with softmax activation. The proposed optimal number of feature maps is (35, 45). The minimum validation error has been found to be 1.22% which is the current best result compared to all other methods in the literature.