waqar-ahmed51/deep-learning-based-noise-type-classification-and-removal-for-drone-image-restoration

Deep Learning-based Noise Type Classification and Removal for Drone Image Restoration

Welcome to the Deep Learning-based Noise Type Classification and Removal for Drone Image Restoration GitHub repository! This repository contains the code and comprehensive documentation for the research paper titled "Deep Learning-based Noise Type Classification and Classification and Removal For Drone Image Restoration".

Abstract

Recent advancements in deep learning have enabled significant progress in image noise type classification and denoising systems. This paper proposes a framework for two-stage multi-type image denoising that provides classification and denoising of four types of noise with a per-class classification accuracy of 98.2% - 100% and a denoising technique that obtained promising PSNR and SSIM values for various types of noise, ensuring effective image restoration. The proposed methodology can be applied to any field that requires image denoising without prior knowledge of the type of noise.

Introduction

This paper proposes a two-stage framework that classifies the noise type followed by denoising based on the predicted noise type. In the first stage, a Convolutional Neural Network (CNN)-based classifier is utilized. The CNN is trained to classify the noisy image into four classes: Gaussian, Salt & Pepper, Poisson, and Speckle. For the second stage of denoising, four separate Denoising Autoencoders (DAEs) have been trained to handle each classified case using its specialized DAE. Lastly, the performance of the proposed framework is evaluated at different stages of the experiment by calculating the peak-signal-to-noise ratio (PSNR).

Proposed Framework Architecture

Datasets

1. CIFAR-100
2. CIFAR-10
3. BSD500
4. Randomly collected drone-captured image samples

Convolutional Neural Network (CNN)

Resources

1. Datasets are available in the repository.
2. Matlab code with instructions is available inside matlab_code_for_dataset_processing for dataset creation.
3. Complete Framework code with instructions is available in image_noise_classification_and_denoising_complete_experiment.ipynb.

Get Started

1. Clone the repository into your local storage.
2. Install Python 3.10.5.
3. Install Jupyter Notebook.
4. Run image_noise_classification_and_denoising_complete_experiment.ipynb each cell one by one by following the instructions.

Note: This repository contains the outcomes of numerous experiments conducted at different stages of our research. Due to the extensive nature of our work, the information in the code may appear scattered. We appreciate your understanding and thank you for your interest in our research.