I have a dataset consisting of characters(lowercase and uppercase) and numbers, totalling about 62 classes. The data I have are about 45 images per class and no test data. The data is a subset of the EMNIST dataset but we aren't allowed to use transfer learning. I am also using data augmentation on the test data. I've used train_test_split with (0.1) test data. I have been experimenting a lot with various CNN architectures but the validation accuracy is coming around 70% with testing accuracy sometimes going up to 90% but I guess it's mostly because of over-fitting.
How do I ensure I am able to make a robust model, or that the model I have made is feasible? Are there some methods I can employ to ensure that the model is able to deal with the small data well?
Also, the images are of size (900,1200) which I have resized to (64,64). Is there a general CNN architecture for this image size that I can use as a base for experimenting?