Face Recognition With OpenCV

In this project we implement face recognition with OpenCV. The first version uses haarcascades as the model to recognize faces. The second version uses the face-recognition module.

Requirements

Version 1

• OpenCV
• Numpy
• Webcam.js (for the front end)
• Flask
• werkzeug.utils
• base64

Version 2

• OpenCV
• Numpy
• Face-recognition

Implementation & Results

Version 1

The first version was implemented using haarcascades.

A front end was developed using Flask to accept input via a webcam, process the image and return a processed image with recognition result.

I had trouble implementing the webcam data capture initially using OpenCV for python. I discovered there was the JS version available as OpenCV.js. However I implemented image capture with Webcam.js and implemented scripts using this tutorial.

The captured image is sent to the Flask app for processing via Ajax XMLHttpRequest(). The POST response contains the processed image filename which the snapshots.html page calls to load and display the processed image.

The faces_train.py script is used to train the model and then save features and labels arrays as .npy files so that they can be reused without having to train the model every time a face recognition task is to be done.

Apart from capturing images via webcam, I provided a file upload method. When the snapshot page is loaded by a client, the webcam doesn't load if the page isn't served via https. The file upload method allows a user to submit an image for face recognition if this app isn't deployed with SSL certificate.

Version 2

The second version uses face-recognition module.

In order to enhance accuracy of face recognition, I sought to implement the project with a more accurate algorithm. The face-recognition module yielded better results.

The faces_trainV2.py script is used to encode faces from training images and save these encoded faces, locations and person names as .npy files which can be reused.

Run the face_recog_videoV2.py file to open your local webcam (or USB webcam plugged into your computer) and run facial recognition on the video feed.

Challenges

1. I had no SSL certificate with which to deploy the app so pages couldn't be served via https. This did not allow client browsers to accept use of webcam by a remote server.
2. Performance of haarcascade models are quite weak. There were frequent wrong recognitions.
3. Because of the Ajax POST call in snapshot.html, render_template in the flask app returned a POST response instead of displaying the expected webpage on the browser. I discovered there appeared to be a difference between an http POST and form POST. For this reason I elected to receive the filename in the POST response and feed that to <img id="result" src="'+sender.responseText+'" width=350 height=300/> in the snapshot.html page. Refer to details of @app.route('/snapshot', methods=['GET','POST'])