Inverse problem in image processing: Image restoration

Wong, Yen Khai (2023) Inverse problem in image processing: Image restoration. Final Year Project, UTAR.

Preview

PDF Download (5Mb) | Preview

Abstract

Conventional image restoration methods often require careful feature selection and fine-tuning, which can be a complicated process and not always possible. On the other hand, Deep-learning (DL) models rely heavily on the datasets availability and neural network architecture, which can lead to reduced performance if the network is poorly designed. Recently, Deep Image Prior (DIP), a learning-free approach to image restoration has emerged as an alternative. However, DIP requires a pre-defined early stopping, which can limit its practical applications. Hence, this project aims to improve image restoration through DIP and address the limitation mentioned. This research proposes the use of Metric-based Early Stopping (MB-ES) with the DIP model for image denoising and super-resolution tasks. The proposed MB-ES algorithm utilizes intermediate restored images to identify the optimal stopping point using PSNR and SSIM metrics, thus eliminating the need for pre-defined early stopping. The results show that MB-ES requires fewer iterations to obtain a better-quality image and has lower design complexity as compared to the existing Early Stopping using Exponential Moving Variance (ES-EMV). The proposed MBES algorithm with DIP is then evaluated on both image denoising and superresolution tasks, and compared with classical and deep learning-based methods. The results show that the proposed MB-ES algorithm achieves remarkable performance in detecting the stopping point that closely resembles the ground truth. In general, the proposed MB-ES on DIP outperforms classical methods and shows comparable performance with recent deep-learning-based models. It is worth noting that DIP does not require heavy training on massive datasets to achieve the performance that DL models possess. The findings of this research are hoped to benefit practical applications especially when dataset is not available and computational resource is limited for DL.

Actions (login required)