Real time Handwritten Digit Identification using Optimized CNN

This paper presents the design and a comprehensive modelling of an efficient Convolutional Neural Network (CNN) for handwritten digit recognition. Experiments we applied our framework to MNIST dataset containing 70,000 gray scale pictures of a digit between 0 and 9. Our designed lightweight CNN architecture of SVMBD contains several convolutional layer, and max pooling layer, followed by few dense connection, and ends with softmax classification layer. Such a setting enables powerful discrimination between several classes with low computational cost. Prior to training the model, we performed important preprocessing to the images, such as scaling and reshaping them to be consistent. We also onehot encoded the label vectors compatible with classification purpose. We chose Adam optimizer with categorical cross-entropy as the loss function for our training, and focused on tuning the most important hyperparameters. We utilized GPU acceleration for performing the computations faster. The dataset was split into 90:10 for training and testing to ensure comprehensive validation. The results were surprising, the model achieved an accuracy of 99.96%. We also found that the critical evaluation measures (precision, recall, and F1-score) are remarkably high for all digit categories. On closer examination of the confusion matrix, we could see that very few misclassified predictions were being made, and that the macro-average ROC-AUC score was perfect, which obviously reflects the model’s superior capability in differentiating the different digits. We did additional tests using new sequences of digits that the model hadn’t previously seen, and that really verified its ability to generalize and do well in the real world. An additional difference is that this CNN architecture is lightweight, and can be employed in mobile and edge computing environments. On the whole, this study creates a stable foundation for developing efficient compact CNN for tackling more sophisticated/relevant image recognition tasks, opening up promising perspectives with regards to automated documents processing, postal classification, high-end visual intelligence systems, etc.