Skip to content

Repository files navigation

Task-Specific-Image-Enhancement


Recovering a clear image solely from an input hazy image is an extremely difficult task to perform. Existing enhancement techniques are empirically expected to be helpful in performing high-level computer vision tasks like object detection. However, it is observed not always to be accurate and practical.

In this project, we have developed three image dehazing methods:
(1) Image dehazing model built using CGAN architecture
(2) Gaussian Blur Image pyramid approach
(3) Color Balancing and Histogram Equalization approach (CBHE approach)
These methods are designed with an aim to merge them with the existing object detector models like Faster RCNNand improve the accuracy of object detection in poor visibility conditions like haze.

Project


  1. cGANs model for Image Dehazing has been implmented on Pytorch
    Prerequisites :
python==3.7
torch==1.2.0
torchvision==0.4.0

Install requirements.txt available in the cGans based approach folder using the following command

pip install -r requirements.txt
  1. Image Pyramid Approach
    Prerequisites:
python==3.7
opencv==4.1.2
  1. CBHE Approach
    Prerequisites:
python==3.7
opencv==4.1.2

Dataset


We are using RESIDE STANDARD Dataset comprising of 13990 indoor hazy images and 1399 clear indoor images.
The dataset is available at: https://sites.google.com/view/reside-dehaze-datasets/reside-standard?authuser=0

Overall Architecture


1) cGANs Approach

cGANs Approach

In the picture above each layer of Generator comprises of a Dense Block(DB)

Loss Function

Loss Function

2) Image Pyramid Approach

Image Pyramid Approach

3) CBHE Approach

CBHE Approach

Authors:

Name github handle
Jatin Dawar @jatin008
Prem Raheja @prem1409
Utkarsh Vashisth @uvashisth

About

Developing an algorithm to enhance the quality of the hazed image using image processing techniques

Resources

Stars

4 stars

Watchers

2 watching

Forks

Releases

Packages

Used by

Contributors

Languages