Experiment - Convolutional neural network (CNN) to recognize errors in textual document formatting

I am looking for an specialist in convolutional neural networks (CNN) to create a three experiments:

First experiment
Create CNN models that recognize specific objects in bachelor thesis:
- Headings – first level heading, subheading and third level heading
- Headings text font
- Headings text size
- Line spacing before and after heading
- Other elements in two groups:
-- Other elements;
-- Footnote references;
- Number in front of the heading

Each object will be recognised by separated CNN model. Each CNN model will have one class – it will be able to recognise only one object.

Second experiment
There will be only one CNN model with class for each object from first experiment.

Third experiment
Create CNN model that can recognize formatting errors in headings.

Expected outcome
-- Clarity on whether CNN can be used to detect errors (or to what level) in the formatting of textual documents.
-- Clarity on exactly how to implement CNN model in such system:
--- Whether to build a CNN model directly for error detection, or to build a CNN model that recognizes elements, therefore, error detection will be done by pre-programmed algorithms.
--- Is a single-class CNN model or a multi-class CNN model works better in situations where CNN model must detect formatting errors in textual documents.

Experiments must be done with Keras or TensorFlow or similar library.
Data sets will be provided for training and for test after request.