05rs / Face-Recognition

End to End Face-Recognition follows the approach described in FaceNet with modifications inspired by the OpenFace project. Semi-hard triplet loss and online semi-hard triplet generator are used for further fine-tuning.

Face-Recognition

End to End Face-Recognition follows the approach described in [1] with modifications inspired by the OpenFace project. Keras is used for implementing the CNN, Dlib and OpenCV for aligning faces on input images. Semi-hard triplet loss and online semi-hard triplet generator are used for further fine-tuning.. Face recognition performance is evaluated on a small subset of the LFW dataset which you can replace with your own custom dataset e.g. with images of your family and friends if you want to further experiment with the notebook.

Technology Used

• Tensoflow
• OpenCv
• Python

Model Architecture

Preprocessing Pipeline

• Directory based image retrival

• Face Extraction using DLib as per FaceNet paper and Face Alignment (eyes and mouth landmark based alignment)

Face Recognition Pipeline

• Used pretrained 'nn4.small2.v1' model [2]

• Creating Siamese Net adding a triplet loss layer.

• Used Tensorflow Semihardloss

• Uses custom semi-hard online triplet Generator for fine-tuning

  • Generate the best triplet in the given batch of training.

Evaluation Pipeline

• Used Support Vector Machine (SVM), K Nearest Neighbour (KNN), Gaussian Naive Bayes (gnb) for inference.

  • Distance Threshold using pretrained model on subset of LFW Dataset.

  

  • Distance Threshold using pretrained model and fine-tuning on subset of LFW Dataset.

  

• t-distributed Stochastic Neighbor Embedding (t-SNE) is applied to the 128-dimensional embedding vectors.

  • t-SNE of pretrained model on subset of LFW Dataset

    

  • t-SNE of pretrained model and fine-tuning on subset of LFW Dataset

    

Status

• Image Extraction and Preprocessing Pipeline complete.

• Model Pipeline complete in Tensorflow.

  • Converted orignal model in csv to binary for tensorflow.
  • Creating Siamese Net adding a triplet loss layer.
  • Used Tensorflow Semihardloss.

• Hard Triplet Generator completed for further Fine-Tuning.

  • Made fast custom Image Data Generator for implementing Semi-Hard online Triplet Generator.

Features

• Online semi-hard triplet generator.

• Fine-tuning to custom dataset.

• Less GPU extensive as uses pre-trained model as base to fine-tune to.

• Fast Generator

Next Todo:

• Visualization
• Model evaluation and metrics