Scribd
Upload
4 views

Frequency-Based Deep Exposure Correction for Low-Light Images

This project focuses on improving low-light image enhancement through a novel frequency-based deep exposure correction method using a CNN U-Net architecture. It aims to address challenges such as noise and detail preservation by incorporating frequency-domain loss functions and optimizing performance for real-world applications. The approach is validated through comprehensive evaluations, demonstrating its effectiveness over traditional methods in various fields including security, autonomous vehicles, and augmented reality.

Uploaded by

Rahul sai
Copyright
© © All Rights Reserved
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
4 views

Frequency-Based Deep Exposure Correction for Low-Light Images

This project focuses on improving low-light image enhancement through a novel frequency-based deep exposure correction method using a CNN U-Net architecture. It aims to address challenges such as noise and detail preservation by incorporating frequency-domain loss functions and optimizing performance for real-world applications. The approach is validated through comprehensive evaluations, demonstrating its effectiveness over traditional methods in various fields including security, autonomous vehicles, and augmented reality.

Uploaded by

Rahul sai
Copyright
© © All Rights Reserved
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
You are on page 1/ 10

Frequency-Based Deep

Exposure Correction for


Low-Light Images
By : Podamala Rahul Sai – 1DT21IS105
Khushi Manjunath – 1DT21IS073
Pooja S K – 1DT21IS108
Preetham S – 1DT21IS114
Introduction
 Low-light images often face challenges like severe
noise and lack of detail.
 Traditional exposure correction methods are limited in
handling these issues effectively.
 There is a growing need for advanced solutions to
enhance image quality in low-light conditions.
 Conventional methods often fail to preserve fine
details and reduce noise effectively.
 This project aims to improve low-light image
enhancement, offering a more robust solution.
 The goal is to develop tools that contribute to higher-
quality images in challenging lighting scenarios.
Literature Survey
1) "Deep Perceptual Image Enhancement Network for
Exposure Restoration" by Panetta et al.
DPIENet is a deep neural network designed to restore images
taken in poor illumination. It uses a structure based on synthesizing
multiple exposures (like creating multiple brightness levels) and
relies on a human-eye-based loss function to evaluate how
perceptually realistic the images look.

Disadvantages:
 Computationally expensive
 Limited generalization to all datasets
2) Frequency-Domain Based Structure Losses for
CycleGAN-Based Cone Beam Computed
Tomography Translation" by Pai et al.
This research focuses on medical imaging, specifically enhancing
cone-beam computed tomography (CBCT) images using
CycleGAN and frequency-based losses to mitigate artifacts.

Disadvantages:
 Narrow application scope:
 Complexity
3)"Deep Photo Enhancer: Unpaired Learning for
Image Enhancement from Photographs with
GANs" by Chen
This paper presents an unpaired learning approach for image enhancement
using a two-way generative adversarial network (GAN) framework. It
enhances images with a Wasserstein GAN (WGAN) and improves
training stability through individual batch normalization layers.

Disadvatages:
 Requires high-quality data
 Dependence on unpaired learning
Problem Statement

The problem of low-light image enhancement arises due


to poor visibility, low contrast, and noise in images
captured under low-light conditions, making it difficult to
extract useful information. Existing methods often
struggle with balancing noise reduction and detail
preservation, and they may require complex setups or
significant computational resources. This project aims to
develop a robust, efficient, and real-time solution that
improves low-light images while maintaining natural
appearance, reducing noise, and enhancing contrast,
suitable for diverse real-world applications.
Abstract

This project presents a novel approach to enhancing low-light


images through deep exposure correction using frequency-domain
analysis. A CNN U-Net-based encoder-decoder model is enhanced
with frequency-domain loss functions, leveraging Fast Fourier
Transform (FFT) and Discrete Cosine Transform (DCT) to
preserve essential high-frequency details and improve
underexposed regions. Comprehensive evaluations using
quantitative and qualitative metrics validate the model’s
effectiveness in surpassing traditional exposure correction
techniques, establishing a robust framework for real-world low-light
image enhancement.
Objectives
 To develop a robust CNN U-Net architecture enhanced with
frequency-domain analysis for low-light exposure correction.
 Incorporate FFT/DCT-based loss functions to retain high-
frequency details while correcting exposure.
 Optimize model performance with curated datasets and hybrid
loss functions for real-world adaptability.
 Evaluate model efficacy through extensive testing on diverse
lighting conditions and benchmark datasets.
 Establish a scalable framework for future research in low-light
image enhancement and restoration.
Applications

 Security and Surveillance


 Autonomous Vehicles
 Aerial Imaging (Drones)
 Forensic Imaging
 Satellite and Space Imaging
 Augmented Reality (AR) and Virtual Reality (VR)
 Military and Defence
THANK YOU

You might also like