Face recognition on OpenCV 3.4.0

The full tutorial can be found on the link: http://www.python36.com/face-recognition-using-opencv-part-1/

Requirements can be found in the file requirements.txt

Files

1> create_database.py (used to create sqlite database)

2> record_face.py (Records the Name associated with the face)

3> trainer.py (used to train LBPHFaceRecognizer for face recognition)

4> detector.py (detects the face from previously trained data and fetches corresponding info from database)

5> train_from_files.py (train using LBPH, eigen, or fisher on any giving dataset as input)

6> test.py (run tests on test data given as input. Outputs the overall confidence score and accuracy of the test data)

Our Contributions

Setup

To use train_from_files.py and test.py, it's required to setup the training and test data directories and files like this:

The training data should have a parent directory for each different person and the pictures of each person should be labeled as username.#.jpg so that each filename is unique.

trainingdata_root/
--> username
----> username.1.jpg
----> username.2.jpg
...
...
----> username.20.jpg
--> anotheruser
----> anotheruser.1.jpg
----> anotheruser.2.jpg
...
...
----> anotheruser.20.jpg

The test data should have all test images in the same parent directory, and labeled similary to the trainingdata

testdata_root/
--> username.17.jpg
--> anotheruser.13.jpg
--> bobby.3.jpg
...

Running the training and tests

1> train_from_files.py Usage:

python train_from_files.py --training-data path/to/trainingdata --learning-algorithm eigen

This is used to train on any training set passed in as a parameter with one of the 3 learning algorithms (eigen, fisher, or lbph). The images are converted to grayscale and resized if needed. A training template is placed in recognizer/trainingData.yml. Test data can be run against this template.

2> test.py Usage:

python test.py --test-data data/testdata --learning-algorithm lbph --error-check

The learning algorithm passed in here should be the same algorithm used in the training template (trainingData.yml). --error-check will calculate the % of tests matched correctly. It's able to determine the correct match by checking the filename in the testdata.

Continuous Authentication

A method to evaluate user authentication continuously was added to detector.py. The video feed will display the confidence score obtained from each prediction and an overall score to determine whether the user is the person they say they are. The 2 scores will update continuously in the camera screen. When the overall score drops below a certain threshold the text and face box will turn from green to red, indicating there may be an issue authenticating the user.