IDBall Readme

Basic Information

1. This repository contains the official implementation of the following paper: "Low-Visibility Scene Enhancement by Isomorphic Dual-Branch Framework With Attention Learning" Paper Link: https://doi.org/10.1007/978-981-99-8073-4_3
2. Authors: Zhan Li; Wenqing Kuang; Bir Bhanu; Yifan Deng; Yihang Chen; Kangkang Xu
3. cite this paper：

@ARTICLE{10879350,
  author={Li, Zhan and Kuang, Wenqing and Bhanu, Bir and Deng, Yifan and Chen, Yihang and Xu, Kangkang},
  journal={IEEE Transactions on Intelligent Transportation Systems}, 
  title={Low-Visibility Scene Enhancement by Isomorphic Dual-Branch Framework With Attention Learning}, 
  year={2025},
  volume={26},
  number={5},
  pages={7127-7141},
  keywords={Transformers;Image enhancement;Visual perception;Superresolution;Image reconstruction;Degradation;Electronic mail;Transportation;Training;Strips;Deep learning;image enhancement;intelligent transportation systems;super resolution;visual perception},
  doi={10.1109/TITS.2025.3534472}}

Abstract

Vision-based intelligent systems are extensively used in autonomous driving, traffic monitoring, and transportation surveillance due to their high performance, low cost, and ease of installation. However, their effectiveness is often com- promised by adverse conditions such as haze, fog, low light, motion blur, and low resolution, leading to reduced visibil- ity and increased safety risks. Additionally, the prevalence of high-definition imaging in embedded and mobile devices presents challenges related to the conflict between large image sizes and limited computing resources. To address these issues and enhance visual perception for intelligent systems operating under adverse conditions, this study proposes an all-in-one isomorphic dual-branch (IDB) framework consisting of two branches with identical structures for different functions, a loss-attention (LA) learning strategy, and feature fusion super-resolution (FFSR) module. The versatile IDB network employs a simple and effective encoder-decoder structure as the backbone for both branches, which can be replaced with task-specific tailored backbones. The plug-in LA strategy differentiates the functions of the two branches, adapting them to various tasks without increasing computational demands during inference. The FFSR module concatenates multi-scale features and restores details progressively in downsampled images, producing outputs with improved visibility, brightness, edge sharpness, and color fidelity. Extensive experimental results demonstrate that the proposed framework outperforms several state-of-the-art methods for image dehazing, low-light enhancement, image deblurring, and super-resolution image reconstruction while maintaining low computational overhead. The associated code is publicly available at https://github.com/lizhangray/IDBall.

Overview

Fig: The IDB framework for low-visibility enhancement of scenes in adverse conditions such as haze, low light, blur, and low resolution.

Our main contributions are summarized as follows.

• A versatile and flexible IDB framework is designed for low-visibility enhancement of diverse images acquired under adverse conditions, including haze, low lightness, blur, and low resolution. The IDB network can process various degraded images to improve visual perception for downstream high-level computer vision tasks.
• A loss attention (LA) learning strategy is proposed for both the pixel and patch levels to distinguish image parts that are “easy” or “hard” to learn. A dynamic mask is automatically computed to differentiate the roles of the two isomorphic branches during the training phase of the IDB network. Additionally, the LA module is detached at inference time to maintain efficiency.
• An SR module for progressive multi-scale feature fusion (PMFFSR) is designed to restore the details in low-visibility LR images. Consequently, large input images are downsampled and then enhanced or recovered with a low computational cost.

Environment

1. Clone Repo

   git clone <code_link>
   cd IDBall-main/
   

2. Create Conda Environment and Install Dependencies

   torch
   tqdm
   piqa
   ...

Prepare models and quick test

1. Prepare your directory structure

   IDBall-main
       |- datasets
           |- ACDC_night
               |- test_gt
               |- test_low
           |- GoPro
               |- test
                   |- GOPR0384_11_00
                       |- blur
                       |- blur_gamma
                       |- sharp
                   ...
           |- SOTS
               |- outdoor
                   |- gt
                   |- hazy          
       |- model
           |- ACDC_night-025.pth
           |- ACDC_night-05.pth
           |- GoPro-025.pth
           |- OTS-025.pth
           |- OTS-05.pth
       |- weights
           |- res2net101_v1b_26w_4s-0812c246.pth
   

2. Download Test Set and Models

   1. Test Set：

      ACDC_night: click here to download

      GoPro: click here to download

      SOTS: click here to download

   2. Models：click here to download

   3. Res2Net weight：click here to download

3. Run eval

   python3 test.py Dehaze-OTS-05
   

   There are four parameters that must be provided:

   • Dehaze-OTS-05: Indicates performing the dehazing task, the model was trained on the OTS dataset, and the input images are 2x downsampled images.
   • Dehaze-OTS-025: Indicates performing the dehazing task, the model was trained on the OTS dataset, and the input images are 4x downsampled images.
   • Delowlight-ACDC_night-05: Indicates performing the low-light enhancement task, the model was trained on the ACDC_night dataset, and the input images are 2x downsampled images.
   • Delowlight-ACDC_night-025: Indicates performing the low-light enhancement task, the model was trained on the ACDC_night dataset, and the input images are 4x downsampled images.
   • Deblur-GoPro-025: Indicates performing the deblurring task, the model was trained on the GoPro dataset, and the input images are 4x downsampled images.

4. The results are saved in the "output" folder.

Visual Examples

To intuitively demonstrate the image enhancement capabilities of the IDB framework across different tasks, we present the following before-and-after image comparisons.

Dehazing

This section showcases the performance of the IDB framework on the dehazing task. Below, you can see a side-by-side comparison of the original hazy image (4x downsampled) and the dehazed image reconstructed by the IDB framework. Observe how the IDB framework effectively removes haze, restoring clarity and color.

Original Image (4x Downsampled)

Reconstructed Image (IDB Framework)

Deblurring

This section demonstrates the effectiveness of the IDB framework in the deblurring task. The comparison below shows the original blurred image (4x downsampled) alongside the deblurred image reconstructed by the IDB framework. Notice the IDB framework effectively reduces motion blur, resulting in sharper image details.

Original Image (4x Downsampled)

Reconstructed Image (IDB Framework)

Low-Light Enhancement

This section illustrates the performance of the IDB framework in the low-light enhancement task. Presented below is a side-by-side view of the original low-light image (4x downsampled) and the enhanced image produced by the IDB framework. See how the IDB framework significantly increases brightness in low-light conditions while preserving details and color.

Original Image (4x Downsampled)

Reconstructed Image (IDB Framework)