An optical character recognition model built on EMNIST for my Arkwright Engineering Scholarship, reaching 82% accuracy on characters outside the training set.
By Niran RadiaPython · TensorFlowSequential CNN
The dataset is more than 700,000 images spanning 62 classes of handwritten character, each one paired with its label.
The report covers the aims, how the data was prepared, the augmentation and scheduling used in training, and what the model still gets wrong.