irisvision 0.1.0

IrisVision - Face Recognition Suite | Made in Emirates 🇦🇪

🇦🇪 IrisVision - Face Recognition Suite

IrisVisionAixL - A high-accuracy face recognition model achieving 98.58% accuracy, designed and trained in Dubai, UAE.

✨ Features

• 🎯 98.58% Accuracy - Beats Buffalo_L benchmark
• 📐 512-dimensional embeddings - Compact and efficient
• 🚀 Easy to use - Simple API for encoding, comparing, and verifying faces
• ☁️ Auto-download - Model weights automatically downloaded from HuggingFace
• 🔧 Production-ready - Optimized for real-world deployment
• 🇦🇪 Made in Emirates - Product of Dubai

📦 Installation

pip install irisvision

🚀 Quick Start

from irisvision import IrisVisionAixL

# Load model (auto-downloads from HuggingFace)
model = IrisVisionAixL.load()

# Encode a face image
embedding = model.encode("face.jpg")  # Returns (512,) numpy array

# Compare two faces
similarity = model.compare(emb1, emb2)  # Returns 0.0 to 1.0

# Verify if same person
is_match, score = model.verify(emb1, emb2, threshold=0.4)

📖 API Reference

IrisVisionAixL.load()

Load the model from HuggingFace Hub.

model = IrisVisionAixL.load(
    repo_id="vCodesUAE/IrisVisionAixL",  # Optional: custom repo
    device="cuda",                        # Optional: "cuda" or "cpu"
    token="hf_xxx"                        # Optional: HuggingFace token for private repos
)

model.encode(image)

Extract face embedding from an image.

# From file path
embedding = model.encode("path/to/face.jpg")

# From PIL Image
from PIL import Image
img = Image.open("face.jpg")
embedding = model.encode(img)

# From numpy array
import numpy as np
img_array = np.array(img)
embedding = model.encode(img_array)

# Batch encoding
embeddings = model.encode([img1, img2, img3])

Returns: numpy.ndarray of shape (512,) or (N, 512) for batch

model.compare(emb1, emb2)

Calculate cosine similarity between two embeddings.

similarity = model.compare(emb1, emb2)
# Returns: float between -1.0 and 1.0 (higher = more similar)

model.verify(emb1, emb2, threshold=0.4)

Verify if two embeddings belong to the same person.

is_match, score = model.verify(emb1, emb2, threshold=0.4)
# Returns: (bool, float)

🏢 Production Example: Attendance System

from irisvision import IrisVisionAixL
import numpy as np

class AttendanceSystem:
    def __init__(self):
        self.model = IrisVisionAixL.load()
        self.employees = {}  # name -> embedding
    
    def register(self, name, face_image):
        """Register a new employee"""
        self.employees[name] = self.model.encode(face_image)
    
    def identify(self, face_image, threshold=0.4):
        """Identify an employee from face"""
        embedding = self.model.encode(face_image)
        
        best_match = None
        best_score = -1
        
        for name, stored_emb in self.employees.items():
            score = self.model.compare(embedding, stored_emb)
            if score > best_score:
                best_score = score
                best_match = name
        
        if best_score >= threshold:
            return best_match, best_score
        return None, best_score

# Usage
system = AttendanceSystem()
system.register("Padam", "padam_face.jpg")
system.register("Ahmed", "ahmed_face.jpg")

name, score = system.identify("unknown_face.jpg")
if name:
    print(f"Identified: {name} (score: {score:.2f})")

🔧 Technical Details

Specification	Value
Architecture	IResNet-100
Configuration	[3, 13, 30, 3]
Embedding Dimension	512
Input Size	112 × 112
Training Loss	ArcFace
Accuracy	98.58%
Parameters	~65M

📊 Model Performance

Metric	Value
Validation Accuracy	98.58%
Classes	405
Embedding Norm	1.0 (L2 normalized)

🔗 Links

• Model Hub: huggingface.co/vCodesUAE/IrisVisionAixL
• GitHub: github.com/vCodesUAE/irisvision
• Organization: vCodesUAE

📄 License

MIT License - see LICENSE file.

🙏 Acknowledgments

• Made with ❤️ in Dubai, UAE 🇦🇪
• Built by vCodesUAE
• Powered by PyTorch and HuggingFace

---

🇦🇪 Made in Emirates | Product of Dubai