search

Crazz-Zaac / computer-vision-experiments

This project delves into the applications of neural networks in image processing and computer vision, driven by our curiosity about their potential beyond theoretical foundations like the Universal Approximation Theorem.
https://crazz-zaac.github.io/computer-vision-experiments/
Apache License 2.0
1 stars 0 forks source link
ai artificial-intelligence computer-vision experiment hacktoberfest image-processing machine-learning neural-network neural-networks opencv unet

readme

cv_expt - Computer Vision Experiments

Dependabot Status GitHub License commit activity code size in bytes Tests Coverage Status

This project is being done with the help of ChatGPT and CoPilot.

Why are we doing it?

We were surprised when learning about the history of Neural Networks and one particular research we liked the most about Neural Network comes from Hornik, Stinchcombe, White 1989 about Existence Theorem (about Universal Approximation Theorem). It states that 3-layer neural nets can approximate any continuous function on a compact domain. Compact domain means input set has a finite diameter and the boundary is part of the input set.

But what about the complex tasks that we try to generalize on real world? And many more...

Installation

  1. git clone https://github.com/Crazz-Zaac/computer-vision-experiments.git
  2. pip install -e cv_expt -r requirements.txt
  3. Tada....

Collaboration

  1. Make an issue and suggest your changes.
  2. We welcome your work. :)

Running Tests

  1. pip install pytest
  2. We recommend to use VSCode and select the tests folder.

Developing

  1. Install black for formatting.
  2. Make sure to enable format on saving for VSCode.

To Do

Order by priority!

Directory Structure

Directory Structure ``` |── data │ └── # Directory for storing raw and processed datasets. ├── setup.py │ └── # Setup script for installing dependencies and setting up the project. ├── cv_expt │ ├── base │ │ ├── defs │ │ │ └── configs.py │ │ │ # Configuration definitions for experiments. │ │ ├── data │ │ │ └── base_dataset.py │ │ │ # Base class for dataset handling and preprocessing. │ │ ├── trainer │ │ │ └── base_trainer.py │ │ │ # Base class for training routines and loops. │ │ ├── logger │ │ │ └── base_logger.py │ │ │ # Base class for logging experiment results - and metrics. │ │ └── models │ │ └── base_model.py │ │ ├── Implements training and inference modes. │ │ └── Contains input/output processing logic for models. │ ├── vis │ │ └── visualization.py │ │ # Visualization utilities for experiment results. │ ├── logger │ │ ├── local_logger.py │ │ │ # Logger for saving logs locally. │ │ └── wandb_logger.py │ │ # Logger for integrating with Weights & Biases (WandB). │ └── experiment │ └── experiment.py │ # Script to define and run experiments. ├── notebooks │ ├── expt1_name.ipynb │ │ # Jupyter notebook for Experiment 1. │ └── expt2_name.ipynb │ # Jupyter notebook for Experiment 2. ├── assets │ └── # Directory for storing images and files used in the README or documentation. ├── models │ └── # Directory to store model weight files for logging and evaluation. ├── outputs │ ├── # Directory for storing local outputs, not committed to the repository. │ └── results │ ├── expt1 │ │ ├── logs │ │ │ └── run_logs.logs │ │ │ # Log files for tracking the progress and results of Experiment 1. │ │ ├── epoch_0.png │ │ │ # Sample output image from the first epoch of Experiment 1. │ │ └── best_model.pth │ │ # The best-performing model checkpoint from Experiment 1. ```

Experiments

Can U-Nets perform Grayscale to RGB? - Please follow [Experiment Notebook](notebooks/grayscale_to_rgb.ipynb) to reproduce results. - **Why is it relevant?** - U-Nets are powerful in semantic segmentation and they are a form of Auto-encoders. Auto-encoders are foundational models for VAE and GANs and can still be good start for an experiment. - **What our experiment showed?** - MSE Loss: ![](assets/gray_to_rgb/metrics_plot.png) - Validation result at Last Epoch: ![](assets/gray_to_rgb/14_0.png) - Test Result on unseen data. ![](assets/gray_to_rgb/test_images.png) - **Conclusions** - It showed some fine results but not strong to say it actually worked. - Surprisingly test results were also not bad. - **Future Works** - Train for longer period? - Use LaB colorspace instead of RGB?
Can U-Nets perform histogram equalization? - **Why is it relevant?** - Histogram equalization can enhance the contrast of images, which might be useful for improving the quality of certain images or features. - **What our experiment showed?** - MSE Loss: ![](assets/hist_eq/metrics_plot.png) - Validation result at Last Epoch: ![](assets/hist_eq/90_2.png) - Test Result on unseen data. ![](assets/hist_eq/test_images.png) - **Conclusions** - Looks like network's prediction is not impressive on test images because it's kind of added artifacts. - **Future Works** - Compare results with CLAHE.
Can U-Nets perform image completion? - **Why is it relevant?** - Image completion, or inpainting, is used to fill in missing parts of images, which can be useful in various applications such as restoration or data augmentation. - **What our experiment showed?** - MSE Loss: ![](assets/img_comp/metrics_plot.png) - Validation result at Last Epoch: ![](assets/img_comp/999_2.png) - Test Result on unseen data. ![](assets/img_comp/test_images.png) - **Conclusions** - This showed that it couldn't generate the missing part of the image, which is quite obvious. U-Net's architecture isn't designed to predict the missing pixels but only for classifying the existing ones. That is, it is best suited for any segementation task.
Can U-Nets perform image super resolution? - **Why is it relevant?** - Image super resolution, simply refers to the method of enlarging image from the smaller size. We can train it on the semi-supervised manner by selcting an original image. Resize to smaller size and pass as an input then compare output with original image. - **What our experiment showed?** - ![]() - **Conclusions** -

Collaborators

--