Research Project

Image Restoration Using Deep Learning

(Image Denoising using Vision Transformer)

Research Team

Lead Researchers:

Xiyi Hang

Collaborators:

Student Team:

Hong Shen
Alex Calderon-Perez
Francisco Hernandez

Funding

Funding Organization:
Funding Program:

SYNOPSIS

Using deep learning techniques to remove noise (visual distortion) from images for clearer, more accurate visuals.
Neural networks are trained to identify and reduce noise without losing essential image details.
Applications in medical imaging, satellite imagery, photography, and more.

Abstract

This project aim to develop new deep learning models for image denoising. The new model will be designed by combining vision transformer, depth-wise convolution, and dynamic convolution. The model will be evaluated by comparison with the state-of-the-art swim-transformer model.

Chart Describing Image Restoration Using Deep Learning alongside original photo with noise and enhanced photo without noise

Research Objectives

To create deep learning models that minimize noise in images while maintaining crucial details.
Ensure that denoising models are generalizable across different types of images

Research Methods/Approach

Implement a transformer architecture tailored for image denoising tasks, leveraging self-attention mechanisms for capturing long-range pixel relationships.
Train the model on large-scale datasets of noisy and clean images, optimizing for image restoration and noise reduction.
Use standard evaluation metrics such as PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index) to benchmark the model’s performance against traditional methods like CNNs and autoencoders.

Research Results and Deliverables

Development of a transformer-based model that significantly reduces noise in images.
Performance evaluation demonstrates better results compared to traditional denoising techniques like Gaussian filters or median filters.

Commercialization Opportunities

Applications: Healthcare, space research, and digital media industries
Key Values: High-quality noise reduction applicable in fields requiring high precision image clarity, such as radiology and satellite data analysis
Potential Customers: Hospitals, research organizations, photographers, and image processing software companies

Research Timeline

Start Date:
End Date: