This was an assignment to perform OCR given several pages of text and their bounding boxes, one set of which was entirely clean, and the other of which was progressively noisier.
-
A description of the assignment can be found in Provided_README.md
-
A description of my solution can be found in report.md
Those reading and judging the code should bear in mind that this was a programming experiment.
All files except this readme form the assignment.
Those modified/created by me were:
- code/system.py
- code/data/model.json.gz
- code/evaluate.py (I modified this to show misclassifications)
- report.md
- Results.xlsx (although this was not part of the submission)
If you're lazy (I am too):
python train.py
python evaluate.py dev
Performance on the unseen test pages (not provided, used for marking) and the provided test pages were as follows:
| Page | % Correct (unseen) | % Correct (provided) | % Correct (baseline*) |
|---|---|---|---|
| Page 1 | 96.3 | 97.6 | 98 |
| Page 2 | 96.6 | 98.7 | 98 |
| Page 3 | 94.1 | 96.5 | 83 |
| Page 4 | 82.9 | 86.8 | 58 |
| Page 5 | 72.1 | 74.0 | 39 |
| Page 6 | 59.5 | 63.8 | 29 |
| Avg. | 83.6 | 86.2 | 67.5 |
*The baseline approach is detailed in Provided_README.md as "the approximate performance that you should expect using a basic nearest neighbour and PCA based approach".
I've found one other solution to this assignment on GitHub from my class:
| Page 1 | Page 2 | Page 3 |
|---|---|---|
 |
 |
 |
| Page 4 | Page 5 | Page 6 |
|---|---|---|
 |
 |
 |