A deep learning model for recognizing English letters and digits
Used dataset
The dataset contains 3410 image files containing handwritten digits[0 - 9] and alaphabets ['A' - 'Z', 'a' - 'z'], a total of 62 classes.
Among the 3410 images, 341 are used for testing and the remaining is used for training
Validation accuracy achieved : 86.51%
Inspection of the incorrect predictions in the validation cases reveal some reasons of Not So High valiadation accuracy
- The model faces difficulties in distinguishing small letters and capital letters. For example c and C, x and X, z and Z ans so on.
- Some letters and digits are very similar looking. For example, l and 1, S and 5, and so on. \
Incorrect predictions in validation data:
Test on images outside the dataset
The model also is not quite capable of recognizing wide range of writing style as seen from the test images outside the dataset