This project is a simplified version of ImageNet. The objective is to design a face recognition software based on CNN. Users enter an image containing human faces, then the program returns an annotated image displaying green zones where faces have been detected.

• Python 3.11.5
• PyTorch 2.1.0
• CUDA 11.8

Below are some instructions on setting up this project locally.

Make sure that Python version 3.11.5 and PyTorch version 2.1.0 are installed on your machine.

1. Clone the repo

   git clone https://github.com/Danh-Lan/OT2-Machine-Learning

2. Prepare train and test data Create two folders train_images and test_images under the root directory. In each of them, create one folder named 0 containing non-face images and another named 1 containing face images. Please contact us to get data.

Four available notebooks dedicated to classifiers' training are:

1. 01Original.ipynb: The base notebook. It loads data from train_images folder, displays images and carries out the training process. You can find an implementation of a very simple neural network architecture with only 2 convolutional and 3 fully-connected layers.
2. 2ImbalancedSampler.ipynb: A notebook designed to take into account the imbalanced dataset. There are more images labeled face than non-face in the training dataset. We will employ an existing data sampler that corrects this imbalance in binary classification before training. Please refer to this repository for further explanation.
3. 3DataAugmentation.ipynb: A notebook in which we apply data augmentation on the fly, including random resizing, cropping and flipping images.
4. 4LRScheduling.ipynb: A notebook in which we inject a scheduling mechanism into the model training cycles.

All trained models are saved under the /models directory as .pth files. These models will serve for the face detection task.

The two following notebooks contain the implementation of a face detector that combines four different approaches:

1. FaceRecognition.ipynb: A detector built from sliding window, image pyramid and Non-maxima suppression.
2. ImgSegmentation.ipynb: The input images are segmented before being scanned by the detector implemented in FaceRecognition.ipynb

Distributed under the MIT License.

• Thanh Lam DANG - [email protected]
• Danh Lan NGUYEN - [email protected]

This project is a part of the Machine Learning & Data Analytics course at INSA Lyon, under the guidance of Prof. Stefan Duffner.