🧠 Face Recognition System from Scratch A lightweight face recognition system built entirely with NumPy — no deep learning frameworks. Implements a two-layer Multilayer Perceptron (MLP) trained from scratch using backpropagation to classify faces by identity. How It Works

Preprocessing — Images are center-cropped (to reduce background noise), resized to 100×100, flattened to a 30,000-dim vector, and standardized. Architecture — A fully connected MLP: input (30,000) → hidden (256, sigmoid) → output (N classes, softmax) Training — Cross-entropy loss minimized via gradient descent with Xavier/Glorot weight initialization over up to 2,000 epochs. Inference — Predicts the identity of a test image along with a confidence score (%).

Tech Stack

Python · NumPy · Pillow · Matplotlib · Google Colab

Dataset Structure

/MyDrive/test/

├── PersonA/ ← training images

├── PersonB/

├── PersonC/

└── test/ ← test images

Features

No PyTorch / TensorFlow — pure NumPy implementation Automatic multi-class support (add folders to add identities) Visual prediction output with confidence score Runs entirely on Google Colab with Drive integration