Edge Deployable Deep Learning Model For Image Analysis

video link:

https://drive.google.com/file/d/19oFZnLerR7BqzS7DQgBRUjD6ly3CpQe3/view?usp=sharing

This repository contains the complete pipeline for an edge-deployable computer vision model designed to analyze images and detect insulator defects. The model is trained to be lightweight and optimized, ensuring it runs efficiently on edge devices like the Nvidia Jetson Nano.

Table of Contents

• Dataset
• Tech Stack
• System Architecture
• Repository Structure
• Execution Workflow & Environments
  • Phase 1: Training & Optimization
  • Phase 2: Edge Deployment
• Quantitative Results
• Visual Results
• How to Run Inference
• Contributors
• License

Tech Stack

• Programming Language: Python
• Deep Learning Frameworks: PyTorch, YOLOv8
• Inference & Optimization: ONNX Runtime, TensorRT
• Computer Vision: OpenCV
• Web App UI: Streamlit
• Edge Hardware: NVIDIA Jetson Nano
• Training Environments: Kaggle, Google Colab

---

System Architecture

The pipeline follows a robust teacher-student knowledge distillation approach, combined with structural pruning and ONNX conversion to guarantee real-time performance on edge devices.

graph TD
    A[IDID Dataset] -->|Train| B[Teacher Model<br>YOLOv8m]
    B -->|Knowledge Distillation| C[Student Model<br>YOLOv8n]
    C -->|Structural Pruning| D[Pruned Student Model]
    D -->|Export| E[ONNX Format]
    E -->|Deploy| F[Edge Inference<br>Nvidia Jetson Nano]
    
    style A fill:#f9f,stroke:#333,stroke-width:2px
    style F fill:#bbf,stroke:#333,stroke-width:2px

Loading

Dataset

link : https://www.kaggle.com/competitions/insulator-defect-detection/data

The model was trained and evaluated on the Insulator Defect Detection Dataset (IDID) and open source Internet Image. The dataset consists of images categorized into three main classes:

• 0: good
• 1: broken
• 2: flashover

Repository Structure

├── data/
│   └── sample_images/       # Sample images for testing the model
├── deployment/              # Edge deployment scripts (Jetson Nano/Local PC)
│   ├── app.py               # Streamlit web UI for defect diagnosis
│   ├── optimized_model.onnx            # Final optimized ONNX model
│   ├── batch_inference.py             # Batch image inference script
│   └── webcam_inference.py  # Real-time webcam inference script
├── docs/                    # Project reports and presentations
├── models/                  # PyTorch model weights (.pt)
│   ├── pruned_model.pt      # Pruned YOLOv8 model
│   ├── student_model.pt     # Distilled YOLOv8 student model
│   └── teacher_model.pt     # Heavy baseline YOLOv8 teacher model
├── training/                # Jupyter Notebooks for the training pipeline
│   ├── preprocessing_teacher.ipynb
│   ├── student_training.ipynb
│   ├── pruning.ipynb
│   └── onnx_conversion.ipynb
├── requirements.txt         # Project dependencies
└── README.md                # Project documentation

---

🚀 Execution Workflow & Environments

The development of this project was split across cloud GPUs (for training) and Edge Devices (for inference).

Phase 1: Training & Optimization (Environment: Kaggle / Google Colab)

All .ipynb files in the training/ folder were executed on cloud platforms (Kaggle/Google Colab) to leverage powerful GPUs. The workflow is strictly sequential:

1. preprocessing_teacher.ipynb: Preprocesses the IDID dataset and trains the heavy baseline teacher model (teacher_model.pt).
2. student_training.ipynb: Applies Knowledge Distillation to transfer knowledge from the heavy teacher to a lightweight student model (student_model.pt).
3. pruning.ipynb: Structurally prunes the student model to reduce its parameter count and physical size while maintaining accuracy (pruned_model.pt).
4. onnx_conversion.ipynb: Converts the final optimized PyTorch model to the ONNX format (optimized_model.onnx) for high-speed inference on edge devices.

Phase 2: Edge Deployment (Environment: Nvidia Jetson Nano / Local PC)

Once the optimized_model.onnx model was generated, it was transferred to the Nvidia Jetson Nano for real-time edge deployment.

Jetson Nano Setup Requirements

Before running the deployment scripts on the Jetson Nano, ensure your device is flashed with the appropriate NVIDIA JetPack SDK.

• Official Nvidia Jetson Nano Developer Kit Setup Guide

Note: The deployment/ folder includes an onnxruntime_gpu-1.11.0-cp38-cp38-linux_aarch64.whl file, which is specifically required for enabling hardware-accelerated ONNX inference on the Jetson's ARM64 architecture.

---

Quantitative Results

Model	Size	mAP50	mAP50-95	Device
YOLOv8m Teacher	49.6 MB	0.985	0.904	GPU Kaggle T4
YOLOv8n Student	6.0 MB	0.986	0.860	GPU Kaggle T4
Pruned Model	4.8 MB	0.943	0.779	Kaggle GPU T4
ONNX Model	9.4 MB	~0.943	~0.779	Jetson Nano

---

Visual Results

Hardware Setup

Sample Inference (Input vs Output)

Input Image	Jetson Nano Output

---

🛠️ How to Run Inference

You can run the deployment scripts on a Jetson Nano or your Local PC (using CPU/GPU).

1. Install Dependencies

For Local PC / Cloud Environment:

pip install -r requirements.txt

For NVIDIA Jetson Nano:

# 1. Install ONNX Runtime using the provided wheel for ARM64
pip install deployment/onnxruntime_gpu-1.11.0-cp38-cp38-linux_aarch64.whl

# 2. Install the lightweight Jetson dependencies
pip install -r requirements_jetson.txt

Note: We recommend using the JetPack pre-installed cv2 (OpenCV) to ensure hardware acceleration is retained on the Jetson Nano.

2. Navigate and Activate Environment

cd deployment
source venv/bin/activate  # Or your appropriate virtual environment activation command

3. Choose your Inference Method:

• Batch Image Processing (Processes images from data/sample_images/):

  python batch_inference.py

• Real-Time Webcam Processing:

  python webcam_inference.py

• Interactive Web UI (Streamlit):

  streamlit run app.py

Note: When you are done, you can exit the virtual environment by running deactivate.

Processed output images and logs will automatically be saved to the deployment/output/ directory.

---

👥 Contributors

Batch No. 9

• Gaurav Chhajed (522261) - GauravChhajed
• Jay Kishan (522136) - Jaysah02
• Abhishek Singh (522101) - singhabhi137

Supervisor:

• Dr. Sri Phani Krishna Karri

---

GitHub Repository: Edge_deployable_DeepLearning_Model_for_Image_Analysis

---

License

This project is licensed under the MIT License.