Deep Learning Project

Exam project for the Deep Learning course @ UniTS

Give it a try!

The quickest way to get started is to train the models using the provided Jupyter Notebooks. You can try them out quickly on Google Colab:

• Denoising Autoencoder trained on MNIST
• Denoising Autoencoder trained on CIFAR10
• Variational Autoencoder trained on MNIST
• Variational Autoencoder trained on CIFAR10
• Denoising Variational Autoencoder trained on MNIST
• Restormer

Otherwise, check out the setup section.

Project description

Image denoising is a fundamental problem in computer vision, with a wide range of real-world applications, including improving the quality of images for object segmentation, detection, tracking, and more. This project aims to explore and implement state-of-the-art deep learning models for image denoising, with a particular focus on Restormer, a Transformer based denoising network.

Deep Learning has demonstrated remarkable success in this field [1], and our project would leverage this technology to address the issue of noisy images. We’d delve into the principles of deep learning, especially Convolutional Neural Networks (CNNs), and understand how these networks can be trained to remove noise and enhance image quality.

One of the key highlights of this project is the exploration of the Restormer architecture, a novel approach to image denoising[2]. The Restormer model efficiently combines the strengths of Transformers and addresses the limitations of CNNs, making it particularly suitable for high-resolution image restoration tasks. By investigating the Restormer model, we aim to understand how it captures long-range pixel interactions while maintaining computational efficiency, thus achieving state-of-the-art results in various image restoration tasks. Our project will focus on exploring the Restormer model for various types of image denoising, including Gaussian grayscale/color denoising and real image denoising. Through this exploration, we anticipate gaining valuable insights into the capabilities of Restormer and its potential applications in practical computer vision scenarios.

We would implement both Restormer and simpler deep learning models for image denoising using Python. After this stage, we would perform an extensive analysis of the results obtained. Evaluation will include qualitative and quantitative metrics. Qualitatively, aspects such as edge preservation, texture, uniformity, and smoothness will be considered. Quantitative metrics like Peak Signal-to-Noise Ratio (PSNR)[3], Structural Similarity Index Measurement (SSIM)[4], and Mean Square Error (MSE) will be used to compare the performance of the deep learning-based denoising methods with traditional techniques. The project will provide a clear understanding of the strengths and limitations of different models.

References

[1] Image Denoising using Deep Learning: Convolutional Neural Network, Shreyasi Ghose et al., 2020

[2] Restormer: Efficient Transformer for High-Resolution Image Restoration, Syed Waqas Zamir, 2022

[3] Peak Signal-to-Noise Ratio as an Image Quality Metric, www.ni.com, National Instruments

[4] Image quality assessment: from error visibility to structural similarity, Zhou Wang et al., 2004

Project setup

To install the GPU version, in order to use the project with a CUDA enabled GPU, run:

conda create --name dl-tf --file conda/tf-gpu.yml

Alternatively, to install the CPU version, run:

conda create --name dl-tf --file conda/tf-cpu.yml

Compile report

The project report is provided as a set of LaTeX source files. The build process is managed by latexmk.

The following packages are required for a successful compilation and can be installed with the following commands on a Linux system using the apt package manager:

apt install texmaker texlive-babel-italian texlive-hyphen-italian texlive-subfigmat texlive-appendix
apt install latexmk

In order to compile the output report/out/main.pdf file, using the provided configuration file, simply run the following commands:

cd report/
latexmk