Generates the data required for training the HMM.

Takes a pixelized image and reconstructs the hidden string

Generates test data and checks it with the already trained model. Returns two values: - Accuracy: Percentage of correctly reconstruct

Reconstruct the string from the HMM results, e.g. [('a', 'b'), ('b', 'c')] -> 'abc'

Takes a flattened list of windows to determine the probability matrices of the hidden markov model Note that the windows have to be c

Returns a HMM object with random probabilities

Calculate the bounding boxes for every character. Source: https://github.com/python-pillow/Pillow/issues/3921

Calculate the overlap between two intervals Example: a=(10, 30) and b=(20, 40) gives an overlap of 10

Pixelize an area of an image, given the parameters

Create an image of size img_size that consists of blocks of size block_size. It is assured that all blocks have a different color

Return a copy of an original image with the character bounding boxes drawn onto it for visualization

Given a list of texts and a font, generate images with that text and font Padding will be added around the text to allow space for pixelizati

Pixelizes the original images with the given block_size. By default, the pixelization is in line with the baseline of the text and the right

Generates n_img strings that follow the given regex pattern

Generates the windows from the pixelized images. Note: The information from the original images is also needed, since we need to infer the ch

Calculate the probability that state X emits symbol Y and save the (row-wise) normalized sum in emission_probabilities[X, Y]

Calculate the probability of starting in state X

From the given windows, count how many times state X follows state Y and save the (row-wise) normalized sum in transition_probabiliti

https://stackoverflow.com/questions/2460177/edit-distance-in-python

Modified edit distance, normalizing the Levenshtein distance between 0 and 1, where 1 indicates a perfect match of the recovered string to th

Takes a list of clustered windows and returns the most likely sequence of characters

Repeating the experiments from Ch. 3.3 of the original publication. It is one of the simplest tasks: Redacted text consists of exactl

Regular viterbi and log-viterbi should return the same values (for shorter sequences)

When the observation sequence gets too long, the regular viterbi will fail due to numerical underflow