Live AI Image Recognition Engine

Overview

The Live AI Image Recognition Engine is a real-time computer vision system designed to perform image classification on live video streams. It leverages deep learning models built with TensorFlow/Keras and integrates with OpenCV for real-time frame capture and inference. The system is optimized for low-latency prediction and can be adapted to various use cases such as face recognition, intoxication detection, and industrial quality control.

---

Features

• Real-time image classification using webcam input
• CNN-based architecture with transfer learning (MobileNetV2 / EfficientNet)
• Support for imbalanced datasets using class weighting or focal loss
• Modular training and inference pipeline
• Performance evaluation with accuracy, precision, recall, F1-score, and confusion matrix
• Optimized deployment using TensorFlow Lite for edge devices
• Scalable design for multiple computer vision applications

---

Tech Stack

• Python
• TensorFlow / Keras
• OpenCV
• NumPy / Pandas / Matplotlib
• TensorFlow Lite (for deployment)

---

System Architecture

Input (Webcam / Image Stream)
        ↓
Preprocessing (Resize, Normalize)
        ↓
CNN Model (Feature Extraction + Classification)
        ↓
Prediction Output (Label + Confidence)
        ↓
Display / Application Layer

---

Project Structure

realtime-image-classifier/
│
├── data/                  # Dataset (train, validation, test)
├── models/                # Saved models (.h5 / .tflite)
├── notebooks/             # Training and experimentation notebooks
├── src/
│   ├── preprocessing/     # Data preprocessing scripts
│   ├── training/          # Model training pipeline
│   ├── evaluation/        # Metrics and evaluation scripts
│   └── inference/         # Real-time inference (OpenCV)
│
├── requirements.txt
├── config.yaml
└── README.md

---

Installation

git clone https://github.com/your-username/live-ai-image-recognition-engine.git
cd live-ai-image-recognition-engine

pip install -r requirements.txt

---

Training

python src/training/train.py

Configure:

• Model backbone (MobileNetV2 / EfficientNet)
• Input size (recommended: 224x224)
• Class weights or focal loss for imbalance handling

---

Evaluation

python src/evaluation/evaluate.py

Metrics include:

• Accuracy
• Precision / Recall
• F1 Score
• Confusion Matrix

---

Real-Time Inference

python src/inference/realtime.py

• Captures webcam input using OpenCV
• Performs frame-by-frame prediction
• Displays class label with confidence score

---

Deployment

For optimized inference:

• Convert model to TensorFlow Lite:

python src/inference/convert_to_tflite.py

• Apply quantization (int8 / float16) for reduced latency and size
• Deploy on edge devices or mobile platforms

---

Use Cases

• Face recognition systems
• Alcohol intoxication detection
• Industrial defect detection
• General object classification

---

Future Improvements

• Integration with FastAPI for web-based inference
• Mobile deployment using TensorFlow Lite
• Temporal smoothing for video stability
• Advanced detection pipelines (e.g., face detection + classification)

---

License

This project is licensed under the Apache License 2.0. See the LICENSE file for details.

---

Author

---

Au Amores

Developed as part of a real-time computer vision system for practical AI deployment and research applications.