This project is divided into 2 tasks:
a) Train two models to recognize handwritten English letters and compare their performances;
b) Use the best trained classifier in the first task to identify which letters compose a corrupted image (captcha)
Task a):
It consists of images of handwritten English alphabet and the corresponding labels. In total 124800 images and labels are present.
For each label 4800 different images are available.
Task b):
It is composed of a series of 4 letters in a corrupted image of size 30 × 140. No test set was provided for this task.
Task a)
K-NN was utilized as a baseline and compared with the classification accuracy of a 2-D CNN.
Task b)
First, noise was removed from captcha. Then, bounding boxes were used to divide the image into 4 letters.
Hence, predictions were made per each letter.
Task a)
| Accuracy Score(%) | Validation set | Test set |
|---|---|---|
| K-NN | 82.9% | 84.8% |
| CNN | 95.01% | 95.4% |