June 18, 2023

A Data-Centric Solution to NonHomogeneous Dehazing via Vision Transformer

�Yangyi Liu¹, Huan Liu¹, Liangyan Li¹, Zijun Wu², Jun Chen¹

¹ McMaster University ² China Telecom Research Institute

mcmaster.ca

mcmaster.ca

|

Introduction

2

NonHomogeneous Dehazing Challenges:

​

• Assumption violation

​

• Limited paired data

We need a solution that:

​

• Leverages the previous years’ data to realize data augmentation

​

• Makes use of the existing knowledge about images in general

​

• Adapts to the target images

mcmaster.ca

|

Methods

3

Data-Centric Engineering – RGB Channel-wise Gamma Correction:

mcmaster.ca

|

Methods

4

Two Branch Framework

​

Transfer Learning Branch – extract pertinent features with pre-trained weights initialization

​

Data Fitting Branch – complement the knowledge learned from the other branch by training from scratch

mcmaster.ca

|

Results

5

Ablation Study & Comparison with SOTA

mcmaster.ca

|

Thank You

​

Q & A

mcmaster.ca

|