Hyper-Spiral Meta-Tuned EfficientNet for Robust Handwritten Digit and Character Recognition Using MNIST and KMNIST Datasets

Handwritten digit and character recognition is among the basic problems in computer vision, given the large variability of human writing patterns, pen thickness, and complexity of images. Though deep learning algorithms demonstrate tremendous performance on standard test problems, namely, MNIST, their performance may remain untrustworthy under varied handwriting patterns of complex characters. To overcome the aforementioned issue, the research proposes the Hyper-Spiral Meta-Tuned EfficientNet approach combined with the Particle Swarm Optimization (PSO) technique for the purpose of reliable recognition of the MNIST and the KMNIST datasets, which include simple Arabic digits and complex Japanese cursive letters, respectively. The Proposed Hyper-Spiral Meta-Tuned EfficientNet makes use of smart pre-processing techniques, PCA for dimensionality reduction, and hyper-spiral PSO with inertia boost capability for hyper parameter selection of EfficientNet. Experimental analysis is done on both the datasets using 80:20 splitting for training as well as testing purposes. The Proposed approach registers 0.9902% accuracy on MNIST and 94.85% accuracy on KMNIST, which is better than the performance of the traditional CNN, SVM, Random Forest algorithm, and baseline EfficientNet models. These analysis results assert that the inclusion of KMNIST has greatly enhanced the robustness, generality, and applicability of the proposed approach for digit and character recognition purposes.