kaushalprasadhial / face-recognition

face-recognition

Face recognition program, capable of recognizing my face as well as 5 other celebrities face. it's a deep learning based program which uses MTCNN for face detection and modified VGG-Face model for face recognition, which is trained on custom data set. I have used VGG-Face model as my base model.
The structure of the VGG-Face model is demonstrated below. Research paper denotes the layer structre as shown below. link to the pre-trained model's weight is here

used transfer learning for the purpose of this assingment. trained the model by keeping the bottom layers unfrozen and other layers frozen

compiled the model by using Adam optimizer with learning rate equal to 1e-5, categorical_crossentropy is used as the loss function and categorical_accuracy as the Evaluation Metric

categorical_accuracy checks to see if the index of the maximal true value is equal to the index of the maximal predicted value.
Categorical crossentropy is a loss function that is used for single label categorization. This is when only one category is applicable for each data point. In other words, an example can belong to one class only.
Use categorical crossentropy in classification problems where only one result can be correct.

Note

This model is trained on Google Colabrotory with GPU enabled runtime. After training for around 400 epochs I got the categorical_accuracy as 70-75%. I have used Google Colabrotory just to train the custom VGG-Face model. and later downloaded that model to my local drive in order to recognize face from webcam.

How to use

First install MTCNN. MTCNN is a deep learning architecture for facial detection. I have used ImageDataGenerator for training the model. For training I had to make a special directory structure in order to use ImageDataGenerator. First there are two directory test and train, and each of them contain 6 sub directories, one for each class of our training and sample. all the sub directories are named as there respective class ('Scarlett Johansson', 'Willa Holland', 'alexandra daddario', 'amber heard face', 'anne hathaway', 'kaushal'). Each sub directory contains only the face of the respective persons. Load the pre-trained VGG-Face model from here. Pretrained model is in .h5 format. Load the model using load_model() Then remove the bottom layers of the VGG-Face model and add your custom layers according to your need. I have added a dense layer, a dropout layer and a softmax layer Train the new custom model with flow_from_directory. I trained the model for around 400 epochs. Train the model up to your satisfaction. First I kept the layers of original VGG-Face model frozen and trained only the newly added layers. Then fine tuned the entire model. You can save the model using save(). Weights for custom model in here Now you have MTCNN to detect the face and custom VGG-Face model to recognize the face. Use OpenCV to read the image. Pass it to MTCNN to detect faces. Pass the detected Faces to our custom VGG_Face model to recognize it.