face_recognition_demo

This is a simple example for face verification using facenet : A unified Embedding for face recognition & clustering implemented by davidsandberg.

Requirements

Linux (Tested on Windows 10)
Python
Python Packages
 - numpy
 - opencv-python
 - TensorFlow
 - scipy
 - sklearn
 - pickle

you can install the python package using pip install <package_name>

Inspiration

I came across FaceNet network, which performs the task of face recognition, verification & clustering. So I decided to try my hands on this fancy stuff as well.

Model Download (pretrained)

Download the MTCNN model from here and place it under mtcnn_model directory.
NOTE: The MTCNN models have already been checked in the repo, so you will see the related files under mtcnn_model/ directory but you are free to change or replace it with your own custom models.
Download facenet model from here and copy to model directory After Download and extracting your directories should look something like this

face_verification_demo
├─ mtcnn_model
|   ├─ all_in_one
│   |   ├─ checkpoint.txt
│   |   ├─ mtcnn-3000000.data-00000-of-00001
│   |   ├─ mtcnn-3000000.index
│   |   └─ mtcnn-3000000.meta
├─ model
|   ├─ 20180402-114759
│   |   ├─ 20180402-114759.pb
│   |   ├─ model-20180402-114759.ckpt-275.data-00000-of-00001
│   |   ├─ model-20180402-114759.ckpt-275.index
│   |   ├─ model-20180402-114759.meta
└─ ...

Preprocessing using MTCNN Face Alignment

Multi Task Cascaded Convolutional Networks performs face detection along with detection landmarks (eyes, nose & mouth endpoints). They provide a very high accuracy in real time. MTCNN leverages cascaded architectures with three stages of carefully designed deep convolutional networks to predict the same. Refer to README.md inside mtcnn_src dir to understand its model architecture.

Therefore MTCNN is used to crop faces from the images, which in turn is fed as input to our FaceNet Network.

Overview of the FaceNet Model

FaceNet is a neural network that learns a mapping from images to a compact Eucledian space where distance correspond to measure of face similarity. Therefore similar face images will have lesser distance between them and vice versa. Once we have the feature vector/ Eucledian embedding per image using the deep convolutional network (FaceNet), tasks such as face recognition, clustering & verification can easily be implemented.

The advantage of using the 128D embedding is that you don't need to retrain you model to recoginize new faces. All you need is a single image for an individual to get that embedding once passed throught the network. That embedding can be stored and used as a reference for any future queries. This makes FaceNet even more powerful.

Triplet Loss

FaceNet uses a distinct loss method called Triplet Loss to calculate loss. Triplet Loss minimises the distance between an anchor and a positive, images that contain same identity, and maximises the distance between the anchor and a negative, images that contain different identities.

Conceptually it means, Faces of same identity should appear closer to each other than faces of another identity.

Implementation

1. Database

Here is what the images directory looks like:

Database has several images of persons, for which we will generate the 128D embedding( or feature vector) and will get the eculedian distance of test image with each of the previously created embedding. If the distance is less than a certain threshold, meaning images are from a same person.

NOTE: To avoid generating the embeddings every time we execute the test_script.py we will dump the embeddings in a pickle file & then reuse it for further references.

If you want to run on your own set of images, simply add your own images to images directory.

In my implementation, I have used tensorflow. Additionally I have written facenet_utils.py that interacts with facenet.py and MTCNNWrapper.py to abstract all interactions with both the networks (MTCNN & FaceNet). It also has utility methods to return the 128D feature vector.

 # create database pickle file, if not exists
    if not os.path.exists('dbfile'):
        encodings = face_util.convert_to_embedding()
        store_data(encodings)

    # read the pickle file
    enc_list = load_data()

Above code will generate the 128D feature vector for all the images in the images directory. if dbfile exists in the same working directory, then it will load the embeddings from there.
NOTE: if you have added images into the images directory, make sure to delete the dbfile and then re-run the script.

Visual Representaion as to what script test_img.py is doing is as follows:

Results

Below are the test results:

rktayal / face_recognition_demo Goto Github PK