Deep learning neural network model built with Tensorflow. Image processing with OpenCV and Numpy. Repository file structure:
.
├── network.py # Deep learning convolutional neural network model
├── config.py # Config file for global framework settings
├── collect_faces.py # Script for collecting face images for model training
├── inspect_dataset.py # Script for overview and inspection of image training sets
├── preprocessing.py # Function for preprocessing of images for training/prediction
├── train.py # Script for training of neural net on images in folder 'training_dataset'
├── train_history.py # Function for plotting of model training
├── predict_folder.py # Script for testing predictions on images in a given directory
├── predict_video.py # Script for video stream face recognition using trained model
│
├── model # Neural net models applied
│ ├── face_detect # CNN for face detection
│ └── this # Developed models from this framework
│
├── test_dataset # Folder for images for testing model prediction
│
└── training_dataset # Training images sorted in folders (subfolder name = label)
├── person1
├── person2
├── person3
├── person4
├── person5
└── person6
The applied network is a convolutional neural network setup by Adrian Rosebroch. The network is a compact variant of the VGGNet network, introduced by Simonyan and Zisserman in their 2014 paper, Very Deep Convolutional Networks for Large Scale Image Recognition. The architectures uses 3×3 convolutional layers stacked on top of each other in increasing depth, reducing volume size by max pooling and with fully-connected layers at the end of the network prior to a softmax classifier.
The only image preprocessing applied, is rescaling to 30 x 30 pixel resolution, converting to grayscale and normalizing grayscale values:
The current framework applies a separate neural net model for detection of faces in the pictures of a webcam video stream. These captured frames of faces are applied for collection of training images. The same algorithm is applied for determining faces for prediction during application of the trained model.
The figure below shows results from training of the above described convolutional neural net on a set of family member images. The current model was trained on a total of 1379 images of 6 different family individuals (classes). Total size of training image dataset was only 5.9 MB.
400 epochs of training with the network took 17 minutes on my Dell laptop. The resulting model size was ~10 MB. The resulting model predicts individual faces well for real-time video stream processing.
An interesting feature, it the ability the model gives for predicting percentage similarities between different family members :)
- Adrian Rosebrock's webpages on OpenCV https://www.pyimagesearch.com
- VGG - Visual Geometry Group at Oxford http://www.robots.ox.ac.uk/~vgg/