Graduation Work Project by Paulo de Oliveira Guedes, from the Electronic Engineering course, at the Universidade Federal de Pernambuco (UFPE).
This work creates a screen detector for the brazilian electronic voting machine, using AlexNet, Resnet50 and a modified AlexNet called Displaynet to predict the 4 corners of the screen:
- Python 3.9+
Install necessary libraries:
$ pip install -r requirements.txt
To label make sure you have images in ascending order (0.jpg, 1.jpg, ... , n.jpg).
In "data manipulation" folder use a csv with the fallowing colunms name;x1;y1;x2;y2;x3;y3;x4;y4, put the dataset directory to be labeling and the number of the fisrt image to be read. Then run:
$ python labeling.py
The fallowing commands can be used:
- mouse_click = put point
- f = fullscreen
- e = erase point
- backspace = go to previus image
- space = go to next image (if have 4 points in the screen)
- t = fine tuning
- w a s d = fine tuning controll
- esq = quit
The training code train.py uses Neptune as a experiment tracking tool, to better understanging read the basics.
To train a network use one of the networks provided, or one of your interest, commenting as needed, as explained in the code, then run:
$ python train.py
The code load.py takes a trained network ("model"), a folder with photos (in "test_data") and predicts the points in the photos, saving the images with the points. It is necessary to change the "resize_vector" depending on the chosen network
Change the quoted variables as needed and run:
$ python load.py
Link to trained nets.
Link to my results on Neptune.
Link to my Monography in portuguese.