Deep Facial Recognition System

Description

This project implements a Siamese Neural Network for efficient image triplet analysis, seamlessly integrating OpenCV for image collection. Three categories of images - positive, negative, and anchor - are collected using OpenCV and preprocessed to standardize dimensions to 100x100 pixels in JPEG format.

Research Paper

Link to Research Paper

Workflow Overview

1. Data Collection: Utilizing OpenCV, 300 images are collected from each category: positive, negative, and anchor.
2. Preprocessing: Images are resized to 100x100 pixels and converted to JPEG format.
3. Dataset Creation: Positive and negative image pairs are combined with anchor images, generating a labeled dataset. Positive pairs are labeled as 1, and negative pairs as 0.
4. Training Data Preparation: The dataset is split into training and testing sets in a 70-30 ratio and batched with a size of 8 for efficient processing.
5. Siamese Model Architecture: A Siamese neural network is constructed with an embedding layer for feature extraction and a custom distance layer for similarity computation.
6. Model Training: The Siamese model is trained on the prepared dataset, optimizing recall and precision metrics for performance evaluation.
7. Real-Time Verification: OpenCV is employed for real-time image verification, showcasing the practical application of the trained model.

Embedding Layer

Siamese Model

Tech Stack

Python

Tensorflow

Open-CV

Usage

1. Clone the Repository:

   git clone https://github.com/MOHINI1403/Deep-Facial-Recognisation-System.git

2. Install Dependencies:

   pip install -r requirements.txt

3. Folder Structure:

   • Data Folder:
     • Contains three sections: positive, negative, and anchor images.
     • Use OpenCV to collect images for the positive and negative sections.
     • Negative images can be downloaded here.
   • Application Images Folder:
     • Contains two folders: input_image and verification_images.
     • verification_images folder should contain around 50 images randomly selected from positive and anchor sets.

4. Virtual Environment:

   • Create a virtual environment named faceid:

     python -m venv faceid

   • Activate the virtual environment:
     • On Windows:

       faceid\Scripts\activate

     • On macOS and Linux:

       source faceid/bin/activate

Contribution

Contributions are welcome! Fork the repository, make your changes, and submit a pull request.

Key Components

• OpenCV Integration: Efficient image collection and real-time verification.
• Siamese Neural Network: Architecture for learning image similarity through feature embeddings.
• Custom Distance Layer: Calculates the L1 distance between embeddings for similarity assessment.
• Training Evaluation: Metrics such as recall and precision used to assess model performance.

Further Assitance :

Try Reading this Documentation :