gtziafas / hebrew-characters-hwr21 Goto Github PK

Mentioned at 7/06 lecture, DOI: 10.1109/ICSESS.2010.5552365

https://tqdm.github.io/

For each window (without discarding ascenders/descenders):

• project horizontally/vertically
• get medians of projections
• draw box centered around the coordinates of the medians (with a size such as 75x75)
• predict/get max probability from that box
• add x coordinate of centerpoint + max probability to array
• identify minima again
• crop from the minima (either with lines or with A*)

• [] Read some papers about char segm
• [] Implement naive contours + centre of gravity method
• [] Build templates
• [] Make smarter cutoffs with prob distributions from templates

habakkuk font + imagemorph for synthetic data generation

we must experiment with the following hyper-params (10-fold):
dropout_rates (in cnn.py):
[0.1, 0.1], [0.1, 0.2], [0.1, 0.25], [0.1, 0.33]
learning rates:
1e-04, 5e-04, 1e-03, 5e-03
weight decays:
0.001, 0.005, 0.01, 0.05, 0.1
batch size:
32, 64, 128

From project root run e.g. python3 -m qumran_seagulls.scripts.train_cnn_monkbrill -e 15 -early 2 -bs 64 -lr 0.001 -wd 0.01 -kfold 10

^{Created from JetBrains using CodeStream}

Signature
def segment_characters(line: array) -> List[array]:

Change the classifier for the sliding window, to classify a given image as either "on-character" or "between-characters"

Use a sliding window, run the classifier at each sliding window position, then plot the probabilities of each character with a different color

^{Created from JetBrains using CodeStream}

From what I've seen, the main advantage of Arivazhagan et al. is that it can handle skewed documents. However, none of the scrolls is skewed. I think the fact that A* also allows going diagonally makes it the better choice

^{Created from JetBrains using CodeStream}

• thing one
• thing two
• thing three

Morph.cy currently performs three different augmentations:

1. affine
2. perspective
3. elastic

See the folder "augmented_font_samples" to see some augmented examples. The original samples are located in the folder "dump"