lightcnn-pytorch 1.2.1

LightCNN for fast, accurate and lightweight face verification

LightCNN

Fast and accurate face recognition model optimized for Central Asian faces. This PyTorch implementation provides an easy-to-use interface for face verification and feature extraction.

Features

• LightCNN-V4 architecture optimized for Central Asian faces
• Fast and lightweight face verification
• Automatic face alignment and preprocessing
• Easy-to-use Python API
• Pre-trained model weights included

Installation

pip install lightcnn-pytorch

Quick Start

from lightcnn_pytorch import LightCNN

# Initialize model (automatically downloads weights)
model = LightCNN()

# Verify two face images
similarity, is_same, confidence = model.verify("person1.jpg", "person2.jpg")
print(f"Similarity score: {similarity:.3f}")
print(f"Same person: {is_same}")
print(f"Confidence: {confidence:.3f}%")

Usage

Face Verification

# Compare two face images
similarity, is_same, confidence = model.verify(image1, image2)

The verify method returns:

• similarity: Cosine similarity score (-1 to 1)
• is_same: Boolean indicating if images are of the same person
• confidence: Confidence score (0-100%)

Feature Extraction

# Get face embedding features
features = model.get_features(image)

Input Formats

• File path (str)
• BGR image array (numpy.ndarray)
• RGB image array will be converted to BGR automatically

Device Selection

# Use specific device
model = LightCNN(device="cuda")  # or "cpu"

Model Details

• Architecture: LightCNN-V4
• Input: 128x128 BGR image
• Preprocessing: (x - 127.5) / 128.0
• Output: 256-dimensional feature vector
• Verification threshold: 0.7