This project is my second attempt at tackling the problem of automatic detection of text regions in Manga and "cleaning" them.
To run this code, you'll need a Python 3.6 environment with NumPy, PyTorch 17, and Opencv-contrib packages.
Command line args are not currently set up, so if you would like to see a pass through of the current code, you must change the variable "image" to reflect path of your test image, then run main.py.
If you would like to retrain the Convnet, you must run convnet.py
The approach which I took was to find likely regions of text (contained in main.py) then getting rid of false-positives using a convolutional neural network (detailed in convnet.py). An additional attempt at removing false positives was made with training an support vector machine (which seemed to be too strict)