Leveraging Transfer Learning for Enhanced Handwritten Digit Recognition using EfficientNet and Fine-Tuned CNN

Traditional handwritten digit recognition algorithms have depended heavily on existing information and customized features. The main deep learning methods used in the handcrafted digit recognition (HDR study are convolutional neural networks (CNN) which has lately yielded revolutionary findings. The existence of distortion, blurring and diverse writing styles make it difficult to recognize and classify numerals effectively. To cope with these limitations, variety of methods have been proposed using CNNs for HDR, but these methods are facing with the problem of low computational speed and low accuracy specially in diverse datasets. Recognition accuracy can still improve because of the continuous rise in handcrafted information together with the availability of ample processing power. Here, we are proposing a model to identify handcrafted numerals by leveraging the capabilities of the Transfer Learning (TL) method. This is accomplished by first refining a pre-trained convolutional neural network model for HDR recognition. To get the best results, the pre-trained CNN model must be fine-tuned just enough once features have been extracted. This paper presents a TL strategy aimed at classifying HDR on the MNIST dataset. We implement our model on Google Colab platform using Python that utilize EfficientNet-B4, an already trained CNN model, to retrieve the features followed by a sequential CNN adjusted through fine-tuning as a classifier. Our suggested transfer learning model has an F-1 score of 0.9927 and an accuracy of 99.27% with a loss of 0.73%.