FIND

FIND: Framework for evaluating Image enhancements on Neural networks inferencing in aDverse environments

---

Overview

The goal of this project is to test and evaluate using image enhancements to improve adverse environments before being inferenced on by Neural Networks (NNs). A general overview of this methodology is shown below.

This framework utilizes PyTorch, Torch Lightning, Torch Vision, Utralytics YOLO, OpenCV, and a fork of muggledy/retinex

Currently Tested

Image Enhancements:

• Histogram Equalization (Original Algorithm)
• Retinex (SSR)

Environments:

• Dark
• Overexposed
• Hazy
• Dark & Rainy

Models:

• ResNet 18, 50, 101
• GoogleNet (Inception v1)
• YOLOv5s, YOLOv5m, YOLOv8n, YOLOv8m
• Vit (ViT-B_16)

Datasets:

• ImageNet
• COCO 2017

Additionally Supported (Not Tested) Functionality

Image Enhancements:

• Retinex
  • MSR
  • MSRCP
  • GIMP
  • MSRCR
  • AMSR

Environments:

• Blur
• Histogram Remapping - very experimental
• Distibution Remapping - very experimental

Models:

• Nearly any model from Torch Vision Model Zoo
• Most YOLO models from Ultralytics YOLO

---

Getting Started

1. Clone the repo

git clone https://github.com/jjsuperpower/FIND.git

2. Setup datasets directory (see datasets/README.md for more details)

3. Create a python virtual environment (not required but recommended).

python3 -m venv venv
echo "export PYTHONPATH=$PYTHONPATH:$(pwd)" >> venv/bin/activate

4. Source the environment and set python path (this must be done every time you open a new terminal)

source venv/bin/activate

5. Get and initialize submodules

git submodule update --init --recursive

6. Install the requirements

pip3 install -r requirements.txt

7. If you plan to use jupyter notebooks run tests, you need to make sure it is installed. If you have not installed it, you can install it with

pip3 install notebook

8. You can test if everything is setup correctly by running an example jupyter notebook.

jupyter notebook ./examples

---

List of files

📦FIND
 ┣ 📂doc        - Photos used for documentation
 ┣ 📂examples   - Nearly all the examples are identical but just test different models
 ┣ 📂retinex    - Created by git submodule init
 ┣ 📂src        - Location source code
 ┃ ┣ 📜coco_ds.py           - Wrapper for COCO dataset
 ┃ ┣ 📜mymodels.py          - Contains torch lightning wrapper, and preprocessing builder for using multiple transforms
 ┃ ┣ 📜myutils.py           - Functions for displaying image data
 ┃ ┣ 📜transforms.py        - Where custom transforms are defined
 ┃ ┗ 📜yolo_wrapper.py      - Wrapper for Ultralytics YOLO
 ┣ 📂testing            - Random stuff, and some ideas not fully explored
 ┃ ┣ 📜cb_remap.ipynb       - Remap an image brightness and contrast to match another images brightness and contrast
 ┃ ┣ 📜gen_poster_img.py    - Used to generate images for poster, see doc/poster.pdf
 ┃ ┗ 📜hist_remap.ipynb     - Remap an image histogram to match another images histogram
 ┣ 📜README.md           - This file    
 ┗ 📜requirements.txt    - Dependencies

How To Cite

@misc{sanderson2023case,
      title={A Case Study of Image Enhancement Algorithms' Effectiveness of Improving Neural Networks' Performance on Adverse Images}, 
      author={Jonathan Sanderson and Syed Rafay Hasan},
      year={2023},
      eprint={2312.09509},
      archivePrefix={arXiv},
      primaryClass={eess.IV}
}