🛡️ BTS-Face-Recognition: One-Shot Learning Analysis

Implementing Siamese Neural Networks for Low-Data Face Recognition

📌 Project Overview

This repository explores the implementation of Siamese Neural Networks (SNN) to solve the challenge of face recognition in "low-data" environments. Unlike traditional CNNs that require thousands of images per class, this project utilizes One-Shot Learning to identify individuals (specifically members of the global phenomenon BTS) by learning a similarity function rather than a direct classification.

This project was developed as part of a formal research study at VIT Bangalore, focusing on the trade-offs between model depth, computational performance, and accuracy in face validation tasks.

---

🔬 The Siamese Architecture

The core of this project is a Siamese Network consisting of two identical sister L1-Distanced subsystems. Instead of classifying an image, the network learns to calculate the "distance" between two feature vectors (embeddings).

• Feature Extraction: Utilizes shared weights across twin Convolutional Neural Networks to ensure consistent embedding logic.
• Architecture Specifics: Each twin network consists of four convolutional layers (64 to 128 filters), Max Pooling, and a final 4096-node Dense layer.
• Distance Layer: Implements an L1 distance function to measure the absolute difference between the twin feature vectors.
• Final Prediction: A Sigmoid activation function outputs a similarity score between 0 and 1.

---

🧪 Technical Methodology

1. One-Shot Learning Strategy

The model is designed to recognize a face from a single exemplary image. This is achieved by:

• Pairwise Training: The dataset is structured into "Positive Pairs" (two images of the same person) and "Negative Pairs" (two images of different people).
• Binary Cross-Entropy Loss: The network is optimized to minimize the error in predicting whether two images are a "match" (1) or "not a match" (0).

2. The Data Pipeline

• Preprocessing: Images are resized to 100x100 pixels and normalized.
• Real-time Face Detection: Integrated with OpenCV for live frame capturing and facial bounding boxes.
• Face Verification: The model compares an "Input Image" against a "Validation Image" folder to verify identity based on a learned distance threshold.

---

📊 Performance Analysis & Diagnostic Insights

A key part of this research was documenting the challenges of training high-precision models on specific facial features.

• Metric Evaluation: The model was tested against 50 validation samples per category.
• The "Low Performance" Aspect: The report identifies that while the model learns general features effectively, Precision and Recall were affected by:
  • Intra-class Variance: Close similarities in facial structure among group members.
  • Orientation Sensitivity: Performance drops when faces are not centered, highlighting the need for a Spatial Transformer Network (STN).
• Learning Insight: The analysis demonstrates how L1-Distance based similarity requires significant data augmentation to become invariant to lighting and background noise.

---

🛠 Tech Stack

• Deep Learning: TensorFlow / Keras
• Computer Vision: OpenCV (Live Tracking)
• Languages: Python (NumPy, Pandas, Matplotlib)
• Environment: Google Colab / Jupyter

---

🚀 Future Research Directions

To improve upon the current "Low Performance" baseline, the following enhancements are proposed:

• Triplet Loss Implementation: Moving from pair-based learning to anchor-positive-negative triplets to create better clusters in the embedding space.
• Transfer Learning: Utilizing the VGG-Face or FaceNet pre-trained weights to initialize the twin CNNs.
• Attention Mechanisms: Incorporating spatial attention to focus the network on key identity-defining landmarks (eyes, nose, mouth).

---

Developed by DHARKIVE-STUDIO