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)