Recently, image super-resolution methods have attained impressive performance by using deep convolutional neural networks (DCNNs). In this work, we describe the algorithmic advances and results obtained by the proposed methods in the image Super-Resolution field. First, we propose a series of Single Image Super-Resolution (SISR) methods for effective and efficient super-resolution tasks. We initially propose a deep feed-forward CNNs method that follows the realistic degradation model by handling the blur kernels of different sizes and different noise levels in an unified network. Next, we propose a deep iterative residual CNNs method that follows the image observation (physical) model by exploiting the powerful image regularization and large-scale optimization techniques with the residual learning. Then, we proposed an efficient deep iterative CNNs method that solves the SR task by cascading the deep residual denoiser networks. Second, we propose SR approaches for the Real-World super-resolution problem. For this purpose, we first propose a deep residual GAN-based SR approach with an adversarial training the network for the pixel-wise supervision of the generated realistic LR/HR pairs. Next, we propose a deep cyclic GAN-based SR method by translating the LR to HR domain and vice versa in an end-to-end manner. After that, we incorporate the learnable adaptive sinusoidal non-linearities into the LR and SR network, whose parameters are optimized during the network training. Finally, we explore the multi image super-resolution (MISR) problems. We first propose a deep star GAN-based SR method by training the network with a single model to super-resolve the LR images for the multiple LR degradation domains. Next, we propose a deep iterative burst SR method that adopts the burst photography pipeline by following the image observation (physical) model. Lastly, we discuss the broader impact, limitations of the current research work, and possible future research dimensions in the image super-resolution field.

Recently, image super-resolution methods have attained impressive performance by using deep convolutional neural networks (DCNNs). In this work, we describe the algorithmic advances and results obtained by the proposed methods in the image Super-Resolution field. First, we propose a series of Single Image Super-Resolution (SISR) methods for effective and efficient super-resolution tasks. We initially propose a deep feed-forward CNNs method that follows the realistic degradation model by handling the blur kernels of different sizes and different noise levels in an unified network. Next, we propose a deep iterative residual CNNs method that follows the image observation (physical) model by exploiting the powerful image regularization and large-scale optimization techniques with the residual learning. Then, we proposed an efficient deep iterative CNNs method that solves the SR task by cascading the deep residual denoiser networks. Second, we propose SR approaches for the Real-World super-resolution problem. For this purpose, we first propose a deep residual GAN-based SR approach with an adversarial training the network for the pixel-wise supervision of the generated realistic LR/HR pairs. Next, we propose a deep cyclic GAN-based SR method by translating the LR to HR domain and vice versa in an end-to-end manner. After that, we incorporate the learnable adaptive sinusoidal non-linearities into the LR and SR network, whose parameters are optimized during the network training. Finally, we explore the multi image super-resolution (MISR) problems. We first propose a deep star GAN-based SR method by training the network with a single model to super-resolve the LR images for the multiple LR degradation domains. Next, we propose a deep iterative burst SR method that adopts the burst photography pipeline by following the image observation (physical) model. Lastly, we discuss the broader impact, limitations of the current research work, and possible future research dimensions in the image super-resolution field.

DEEP CONVOLUTIONAL NEURAL NETWORKS FOR IMAGE SUPER-RESOLUTION / Rao Muhammad Umer , 2022 Mar 09. 33. ciclo, Anno Accademico 2020/2021.

DEEP CONVOLUTIONAL NEURAL NETWORKS FOR IMAGE SUPER-RESOLUTION

UMER, RAO MUHAMMAD
2022-03-09

Abstract

Recently, image super-resolution methods have attained impressive performance by using deep convolutional neural networks (DCNNs). In this work, we describe the algorithmic advances and results obtained by the proposed methods in the image Super-Resolution field. First, we propose a series of Single Image Super-Resolution (SISR) methods for effective and efficient super-resolution tasks. We initially propose a deep feed-forward CNNs method that follows the realistic degradation model by handling the blur kernels of different sizes and different noise levels in an unified network. Next, we propose a deep iterative residual CNNs method that follows the image observation (physical) model by exploiting the powerful image regularization and large-scale optimization techniques with the residual learning. Then, we proposed an efficient deep iterative CNNs method that solves the SR task by cascading the deep residual denoiser networks. Second, we propose SR approaches for the Real-World super-resolution problem. For this purpose, we first propose a deep residual GAN-based SR approach with an adversarial training the network for the pixel-wise supervision of the generated realistic LR/HR pairs. Next, we propose a deep cyclic GAN-based SR method by translating the LR to HR domain and vice versa in an end-to-end manner. After that, we incorporate the learnable adaptive sinusoidal non-linearities into the LR and SR network, whose parameters are optimized during the network training. Finally, we explore the multi image super-resolution (MISR) problems. We first propose a deep star GAN-based SR method by training the network with a single model to super-resolve the LR images for the multiple LR degradation domains. Next, we propose a deep iterative burst SR method that adopts the burst photography pipeline by following the image observation (physical) model. Lastly, we discuss the broader impact, limitations of the current research work, and possible future research dimensions in the image super-resolution field.
9-mar-2022
Recently, image super-resolution methods have attained impressive performance by using deep convolutional neural networks (DCNNs). In this work, we describe the algorithmic advances and results obtained by the proposed methods in the image Super-Resolution field. First, we propose a series of Single Image Super-Resolution (SISR) methods for effective and efficient super-resolution tasks. We initially propose a deep feed-forward CNNs method that follows the realistic degradation model by handling the blur kernels of different sizes and different noise levels in an unified network. Next, we propose a deep iterative residual CNNs method that follows the image observation (physical) model by exploiting the powerful image regularization and large-scale optimization techniques with the residual learning. Then, we proposed an efficient deep iterative CNNs method that solves the SR task by cascading the deep residual denoiser networks. Second, we propose SR approaches for the Real-World super-resolution problem. For this purpose, we first propose a deep residual GAN-based SR approach with an adversarial training the network for the pixel-wise supervision of the generated realistic LR/HR pairs. Next, we propose a deep cyclic GAN-based SR method by translating the LR to HR domain and vice versa in an end-to-end manner. After that, we incorporate the learnable adaptive sinusoidal non-linearities into the LR and SR network, whose parameters are optimized during the network training. Finally, we explore the multi image super-resolution (MISR) problems. We first propose a deep star GAN-based SR method by training the network with a single model to super-resolve the LR images for the multiple LR degradation domains. Next, we propose a deep iterative burst SR method that adopts the burst photography pipeline by following the image observation (physical) model. Lastly, we discuss the broader impact, limitations of the current research work, and possible future research dimensions in the image super-resolution field.
Super-Resolution; Deep Learning; CNNs; Restoration; Optimization
Super-Resolution; Deep Learning; CNNs; Restoration; Optimization
DEEP CONVOLUTIONAL NEURAL NETWORKS FOR IMAGE SUPER-RESOLUTION / Rao Muhammad Umer , 2022 Mar 09. 33. ciclo, Anno Accademico 2020/2021.
File in questo prodotto:
File Dimensione Formato  
phd_thesis_final_version.pdf

accesso aperto

Descrizione: Tesi definitiva
Licenza: Creative commons
Dimensione 38.28 MB
Formato Adobe PDF
38.28 MB Adobe PDF Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11390/1224262
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact