freefly518 / face-recognition

Face Detection using open-cv and Face Recognition using Siamese networks using Keras and TensorFlow

Face-Recognition

Face Detection using open-cv and Face Recognition using Siamese networks using Keras and TensorFlow

This project can be used to train a Siamese network for Face Recognition based on either Contrastive Loss and Triplet Loss as well. I have implemented this project using Keras (with TensorFlow backend) and open-cv in Python 3.6.

The complete pipeline for training the network is as follows:

1. Extract faces from human images, which are stored in separate person-wise directories. You can use face_detection.py for this purpose. This python script takes two inputs,

   (i) -n/--new_image-dir which stores the faces of all people in seperate person-wise directories, and

   (ii) -o/--orig_image_dir which is a list of original directory names. Keep in mind that all directories mentioned after -o flag must contain images of all people in seperate folders. Sample call is like

                  python face_recognition.py -n face_images -o people_dir_1 people_dir_2
   

2. Having face images stored in the directory mentioned above, you can now run either face_recog_contrastive.py or face_recog_triplet.py based on whether Contrastive Loss or Triplet Loss is to be used to train the network. Both python scripts takes same inputs, namely

   (i) -b/--batch_size represents the batch-size of inputs to be given to network, it defaults to 16.

   (ii) -d/--img_dir represents directory which was mentioned as target directory above containing face images.

   (iii) -m/--model represents the TensorFlow model which is to be used as base model for Siamese networks. Allowed values are - mobilenet(MobileNetV2), inception(InceptionV3), vgg(VGG19), xception(Xception), resnet(ResNet50) and inception-resnet (Inception-ResNet), defaults to mobilenet.

   (iv) -o/--optim represents the optimizer to be used to train the network. Allowed values are sgd (default), rmsprop and adam. All optimizers are used with default parameters.

   (v) -e/--epochs represents the number of epochs for which the model is to be trained, defaults to 5.

   Sample call is like

                python face_recog_contrastive.py -b 32 -d face_dirs -m resnet -o rmsprop -e 20
   

3. To generate predictions from your trained and saved model, you can use test_predictions.py. This python script takes following input parameters

   (i) -d/--test_dir represents the directory where you have test set of images saved. Keep in mind the directory structure should be same as we created for training set, i.e. person-wise directories of images.

   (ii) -t/--target_dir represents directory where predicted results will be saved.

   (iii) -m/--model_file represents saved model file which will be used for making predictions.

   (iv) -l/--loss represents which model will be used, used to determine the model input, contrastive or triplet

   (v) -n/--num_images represents number of image pairs you want to test accuracy on

   Sample call is like

    python test_predictions.py -d test_dir -t results_dir -m contrastive_resnet_model.h5 -l contrastive -n 10