Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Feature Fusion Attention Network with CycleGAN for Image Dehazing, De-Snowing and De-Raining

Akshat Jain

TL;DR

The paper tackles weather-induced image degradation (haze, rain, snow) under severe data scarcity by integrating Feature Fusion Attention (FFA) networks with CycleGAN. This hybrid approach leverages supervised learning from limited clean samples and unsupervised domain adaptation via unpaired data, achieving notable improvements in PSNR and SSIM on RESIDE and DenseHaze datasets. It provides a detailed implementation including building blocks, dataset loaders, and a web interface for real-time evaluation, highlighting strong performance and practicality in real-world scenarios. The work demonstrates robust restoration across multiple weather conditions with relatively data-efficient training and offers a framework for future enhancements through newer restoration paradigms and interpretable metrics.

Abstract

This paper presents a novel approach to image dehazing by combining Feature Fusion Attention (FFA) networks with CycleGAN architecture. Our method leverages both supervised and unsupervised learning techniques to effectively remove haze from images while preserving crucial image details. The proposed hybrid architecture demonstrates significant improvements in image quality metrics, achieving superior PSNR and SSIM scores compared to traditional dehazing methods. Through extensive experimentation on the RESIDE and DenseHaze CVPR 2019 dataset, we show that our approach effectively handles both synthetic and real-world hazy images. CycleGAN handles the unpaired nature of hazy and clean images effectively, enabling the model to learn mappings even without paired data.

Feature Fusion Attention Network with CycleGAN for Image Dehazing, De-Snowing and De-Raining

TL;DR

The paper tackles weather-induced image degradation (haze, rain, snow) under severe data scarcity by integrating Feature Fusion Attention (FFA) networks with CycleGAN. This hybrid approach leverages supervised learning from limited clean samples and unsupervised domain adaptation via unpaired data, achieving notable improvements in PSNR and SSIM on RESIDE and DenseHaze datasets. It provides a detailed implementation including building blocks, dataset loaders, and a web interface for real-time evaluation, highlighting strong performance and practicality in real-world scenarios. The work demonstrates robust restoration across multiple weather conditions with relatively data-efficient training and offers a framework for future enhancements through newer restoration paradigms and interpretable metrics.

Abstract

This paper presents a novel approach to image dehazing by combining Feature Fusion Attention (FFA) networks with CycleGAN architecture. Our method leverages both supervised and unsupervised learning techniques to effectively remove haze from images while preserving crucial image details. The proposed hybrid architecture demonstrates significant improvements in image quality metrics, achieving superior PSNR and SSIM scores compared to traditional dehazing methods. Through extensive experimentation on the RESIDE and DenseHaze CVPR 2019 dataset, we show that our approach effectively handles both synthetic and real-world hazy images. CycleGAN handles the unpaired nature of hazy and clean images effectively, enabling the model to learn mappings even without paired data.

Paper Structure

This paper contains 45 sections, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Problem Statement
  3. Contributions
  4. Workflow
  5. Related Work
  6. Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks[1]
  7. DehazeNet: An End-to-End System for Single Image Haze Removal [2]
  8. Deep Joint Rain Detection and Removal from a Single Image [3]
  9. Methods and Approaches
  10. Pre-Review Phase Work
  11. Post-Review Phase Work
  12. Model and its functions and Methods utilized
  13. Imports and Initial Setup
  14. Defining Basic Building Blocks
  15. Constructing Residual Blocks and Groups
  16. ...and 30 more sections

Figures (4)

  • Figure 1: Performance of the different dehazing model. Adapted from A. Hu et al., “Unsupervised haze removal for high-resolution optical remote-sensing images based on improved generative adversarial networks,” Remote Sensing, vol. 12, p. 4162, 2020. [Online]. [8]
  • Figure 2: Comparison of Rainy Hazy and Clean Images Processed by the FFA + CycleGAN Model
  • Figure 3: Comparison of Snowy Hazy and Clean Images Processed by the FFA + CycleGAN Model
  • Figure 4: Comparison of Foggy Hazy and Clean Images Processed by the FFA + CycleGAN Model