Handwriting Recognition: Unravelling Performance Diversity and Practical Implications

Offline handwritten text recognition (HTR) helps recognise old documents, students’ study materials, postal codes for letter delivery, admission forms or OMR recognition, recognition of medical prescriptions, and many more. Different machine-learning methods have traditionally used the MNIST dataset as a standard to assess their effectiveness in recognising handwritten digits. This work aims to thoroughly analyse the diversity in performance among several machine learning approaches when applied to the MNIST dataset. This study employs both old and new machine learning algorithms, such as "K-Nearest Neighbours" (KNN), "convolutional neural networks" (CNN), "Decision Tree" (DT), "Random Forest" (RF), and "Linear Regression" (LR). In this research paper, the first step is to train and test each technique on the MNIST dataset and compare the results to standard evaluation criteria such as accuracy, precision, recall, FAR, FRR, data loss, and F1-score. In addition, this paper investigates each technique’s computational efficiency to better understand its practicality. The primary research compares machine learning methods for handwritten digit recognition, concentrating on model optimisation. This study aims to fill the research gap by determining the most effective technique for classifying handwritten images using the MNIST dataset, which includes 70,000 records of 28x28 pixel greyscale images from 10 classes. The results of our study show that CNN outperformed the best with a recall, F1-score, accuracy of 0.9871, and the lowest RMSE of 0.5110. A test loss of 0.0447 demonstrates that CNNs generalise well to unseen data. KNN are reliable and generalisable, with a recall and F1-score of 0.9688, prediction errors (RMSE of 0.7925), and a test loss of 0.3615. The recall of RF resulted in an F1-score of 0.9668. Furthermore, based on the abovementioned methods, the ‘DT’ displayed the lowest performance with an accuracy of just 0.8731 and the most significant prediction mistakes with RMSE of 1.4608. ‘LR’ showed the highest amount of test loss at 0.3745. This provides considerable direction for researchers and practitioners seeking the most appropriate methodology for similar jobs. However, this research does not optimise various hyperparameters for the best-tuned image classification.