Face Recognition with VGG_Face-net transfer learning.

Step 1: Dataset acquisition

Train data

Gathered 50 images of 5 most powerful world leaders Trump,Putin,Jinping,Merkel and Modi of 10 images each. Also my 10 images.

Test data

18 images with 3 images each of above 6 persons including me.

Step 2: Detect faces.

Using dlib cnn face detector find faces in image and crop faces and store them in separate folders sorting by individual person.
Download 'mmod_human_face_detector' to use as 'dlib cnn face detector',br>

! wget http://dlib.net/files/mmod_human_face_detector.dat.bz2 
!bzip2 -dk mmod_human_face_detector.dat.bz2

After extracting faces and storing them in corresponding folder the directory structure will be :

Directory structure :
|Images /
|  |-- (60 images)
|Images_crop /
|  |--angelamerkel
|     |--(10 images)
|  |--jinping / 
|     |--(10 images)
|  |--lakshminarayana / 
|         |--(10 imgaes)
|  |--modi / (10 images)
|  |--putin / (10 images) 
|  |--trump / (10 images)
|Images_test / 
|  |-- .. / (18 images)
|Images_test_crop / 
|  |--angelamerkel / (3 images)
|  |--jinping / (3 images)
|  |--lakshminarayana / (3 imgaes)
|  |--modi / (3 images)
|  |--putin / (3 images) 
|Face_Recognition.ipynb
|mmod_human_face_detector.dat

Step 3: Download and load VGG_face weights.

As vgg-face weights are not available as .h5 file to download,from this article
Download Vgg-face weights from google drive with

! gdown https://drive.google.com/uc?id=1CPSeum3HpopfomUEK1gybeuIVoeJT_Eo

Define vgg-face model architecture in tensorflow-keras and load weights.

Step 4: Get embeddings for faces.

To get embeddings from Vgg-face net remove last softmax layer and it outputs 2622 units in last flatten layer which is our embeddings for each face.
These embeddings are used later to train a softmax regressor to classify the person in image.

Step 5: Train Softmax regressor for 6 person classification from embeddings.

Prepare train data and test data from those 2622 embeddings and feed into a simple softmax regressor with 3 layers containing first layer with 100 units and tanh activation function , second layer with 10 units and tanh activation function and third layer with 6 units for each person with softmax activation.

Predictions

For an image(may contain multiple faces) extract each face,get embeddings,get prediction from classifier network,make bounding box around face and write person name.

My image