Image - DeHaze

Posted on Sep 24, 2023

Project Overview

The AI-ML Based Intelligent De-Smoking/De-Hazing Algorithm is an advanced real-time image and video processing solution designed to enhance visibility in fire-prone environments. Developed as part of Smart India Hackathon (SIH) 2023, this project focuses on improving rescue operations by providing clear visuals of areas affected by smoke and haze, particularly in indoor fire hazards.

Problem Statement (SIH1417)

Fire hazards create dense smoke that significantly reduces visibility, making rescue operations more difficult. Existing dehazing algorithms suffer from high latency and may not perform well in real-time conditions. Our project aims to develop an AI-powered low-latency, high-performance dehazing algorithm to address this challenge effectively.

Key Objectives

  • Real-time dehazing for images and videos
  • Enhancing existing state-of-the-art models
  • Reducing model latency while maintaining accuracy
  • Proposing novel model architectures

Our Solution

We implemented two major improvements over existing dehazing models:

GridDehazeNet Enhancement

  • Integrated Pixel Attention Mechanism to improve dehazing quality
  • Improved SSIM score to 0.9841 (original: 0.9836)
  • Reduced latency to 9.91 ms (original: 9.905 ms)

MixDehazeNet Optimization

  • Introduced Mix Structure Block
  • Incorporated Dark Channel Prior, Wiener Filter, and Sparse Attention
  • Combined with Object Detection for better scene understanding

Technical Approach

  • Datasets Used:
    • Training: RESIDE, HAZE 4K
    • Testing: SOTS (500 indoor images)
  • Tech Stack:
    • Backend: Python, Node.js, Express
    • Frontend: React.js

Real-World Applications

  • Enhancing visibility in fire rescue operations
  • Real-time dehazing in vehicles and aircraft
  • Search and rescue missions in low-visibility conditions

Future Enhancements

🔹 Integration of Image Segmentation
🔹 Further latency reduction for real-time applications
🔹 Deployment as a scalable cloud-based API
🔹 Combining MixDehazeNet with YOLOv5 for enhanced object detection in hazy conditions

This project represents a step forward in AI-powered real-time image enhancement, with direct applications in safety, transportation, and disaster management.

📂 GitHub Repository: SIH2023-PixelEncoders