RESIDE-unpaired

Citation Author(s):
Shiyu
Zhao
Submitted by:
Rong Chen
Last updated:
Mon, 01/22/2024 - 07:23
DOI:
10.21227/66v0-6j78
Data Format:
Research Article Link:
License:
297 Views
Categories:
Keywords:
0
0 ratings - Please login to submit your rating.

Abstract 

With the fast growth of deep learning, trainable frameworks have been presented to restore hazy images. However, the capability of most existing learning-based methods is limited since the parameters learned in an end-

to-end manner are difficult to generalize to the haze or foggy images captured in the real world. Another challenge of extending data-driven models into image dehazing is collecting a large number of hazy and haze-free image pairs for the same scenes, which is impractical. To address these issues, we explore unsupervised single-image dehazing and propose a self-guided generative adversarial network (GAN) based on the dual relationship between dehazing and Retinex. Specifically, we carry out image dehazing as illumination-reflectance separation using a decomposition net in the generator. Then, a guide module is applied to encourage local structure preservation and realistic reflectance generation. In addition, we integrate the model with the outdoor heavy-duty pan-tilt-zoom (PTZ) camera to implement dynamic object detection in hazy environment. We comprehensively evaluate the proposed GAN with both synthetic and real-world scenes. The quantitative and qualitative results demonstrate the effectiveness and robustness of our model in handling unseen hazy images with varying visual properties. 

Instructions: 

Unzip RESIDE-unpaired.zip in the folder <RefineDNet_root>/datasets. Your directory tree should look like:

<RefineDNet_root>
├── datasets
│   ├
│   ├── RESIDE-unpaired
│   │   ├── trainA
│   │   └── trainB
│   ...
...

Comments

hii

Submitted by dattatreya halyal on Sun, 05/05/2024 - 01:03