Benchmark of deep learning models for single image super-resolution (SISR)

Soufi, Omar; Aarab, Zineb; Belouadha, Fatima-Zahra

doi:10.1109/iraset52964.2022.9738274

Cited by 7 publications

(5 citation statements)

References 38 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We trained, tested, and evaluated the proposed image super‐resolution model on five distinct real ground‐truth image datasets and two clinical X‐ray image datasets in this study. Using images from the Div2K [19] dataset, the proposed model was trained using ground‐truth images. The dataset consisted of one thousand high‐quality 2k images, of which one hundred were used for validation, eight hundred for training, and one hundred for testing the proposed model.…”

Section: Resultsmentioning

confidence: 99%

See 1 more Smart Citation

X‐Ray medical image super‐resolution via self‐organization neural networks and geometric directional gradient

Ahmadian

Reza-Alikhani

2022

IET Image Processing

View full text Add to dashboard Cite

Nowadays, due to the rapid development of the Internet and communication networks, remote disease diagnosis by physicians utilizing medical images such as X-rays have increased. However, the transmitted images lose their true quality and resolutions due to several factors, including noise and image compression techniques used to conserve memory space. This paper presents a novel clinical X-ray image enhancement method based on self-organizing neural networks, the ability of first-order derivative operators to detect edges, and improved classification methods by enhancing image features and extending the nearest neighbor algorithms selection of the best similarity feature vector among all sample feature vectors. The proposed method employs first-order image derivation operators to extract edge details and features from images. The feature vectors were preprocessed and segmented into various classes using self-organizing neural networks. The best matching and most similar vectors among the feature vectors are retrieved via the nearest neighbor algorithm. In contrast to conventional and contemporary image super-resolution techniques, the proposed method does not rely on the cost function. It is independent of error backpropagation methods because the competitive nature of the neural network is used to update the weight vectors rather than the gradient descent method. In addition, the performance and accuracy of the proposed image super-resolution method were improved by increasing the number of feature components in the feature vectors through the use of the two-dimensional image gradient. Furthermore, numerous image quality measurement criteria, such as structural similarity index (SSIM) and peak signalto-noise ratio (PSNR), were used to compare the reconstructed images of the proposed algorithm with those of other image enhancement algorithms. The results show that the proposed image enhancement algorithms' high-resolution and high-quality reconstructed image significantly outperformed many other medical image enhancement algorithms in reconstructing the edges and boundaries of objects with a smooth slope, as deemed by expert personnel and compared with image quality measurement criteria.

show abstract

Section: Resultsmentioning

confidence: 99%

“…Using images from the Div2K [19] dataset, the proposed model was trained using ground-truth images. The dataset consisted of one thousand high-quality 2k images, of which one hundred were used for validation, eight hundred for training, and one hundred for testing the proposed model.…”

Section: Database Preparation and Configurationmentioning

confidence: 99%

X‐Ray medical image super‐resolution via self‐organization neural networks and geometric directional gradient

Ahmadian

Reza-Alikhani

2022

IET Image Processing

View full text Add to dashboard Cite

show abstract

“…Algorithm/ Methodology Advantages Disadvantages [4] SRGAN Recover the finer texture details computational complexity [5] Swin Transformer blocks (RSTB) Several swin with better accuracy Reduces memory [6] Swin Transformer On different tasks Swinger beats the state-ofthe-art methods by 0.14 to 0.45dBlevel…”

Section: Referencementioning

confidence: 99%

Enhancing Upscaled Image Resolution Using Hybrid Generative Adversarial Network-Enabled Frameworks

Geetha,

Jebamalar,

Shiney

et al. 2024

IEEE Access

View full text Add to dashboard Cite

Upscaling images typically depends on facial features, such as facial geometry details or references, to rebuild reasonable details. The low-quality input images cannot offer accurate geometric details, and the high-quality details are obscure, limiting performance in practical conditions. This work addresses the problem using an enhanced version of the generative adversarial network (GAN) model that uses rich and varied facial features incorporated into the pretrained StyleGAN2 for face restoration. The generated facial features are integrated into the facial restoration process using spatial feature transform layers to achieve facial details and color quality to improve reliability. In particular, the image background is upsampled using a modified enhanced super-resolution GAN trained in parallel to remove noise while recreating a high-resolution image from low-quality input. The upscale image resolution GAN is used to enlarge the image for a clear view without sacrificing the original image quality. Finally, by combining the upscaled background and restored faces, the hybrid GAN-enabled framework can obtain high-resolution upscaled images. An experimental comparison has been conducted using the Flickr-Faces-HQ Dataset collected from Kaggle. The findings indicate that the suggested framework outperforms existing methods in terms of reconstruction, adversarial, and facial component loss metrics as well as similarity indexes.

show abstract

“…A detailed study of the latest techniques used for SISR can be found in [26]. Here, 53 different models were evaluated on 7 different datasets.…”

Section: Related Workmentioning

confidence: 99%

Single Image Super Resolution Using Deep Residual Learning

Hassan,

Illanko,

Fernando

2024

View full text Add to dashboard Cite

Single Image Super Resolution (SSIR) is an intriguing research topic in computer vision where the goal is to create high-resolution images from low-resolution ones using innovative techniques. SSIR has numerous applications in fields such as medical/satellite imaging, remote target identification and autonomous vehicles. Compared to interpolation based traditional approaches, deep learning techniques have recently gained attention in SISR due to their superior performance and computational efficiency. This article proposes an Autoencoder based Deep Learning Model for SSIR. The down-sampling part of the Autoencoder mainly uses 3 by 3 convolution and has no subsampling layers. The up-sampling part uses transpose convolution and residual connections from the down sampling part. The model is trained using a subset of the VILRC ImageNet database as well as the RealSR database. Quantitative metrics such as PSNR and SSIM are found to be as high as 76.06 and 0.93 in our testing. We also used qualitative measures such as perceptual quality.

show abstract

Benchmark of deep learning models for single image super-resolution (SISR)

Cited by 7 publications

References 38 publications

X‐Ray medical image super‐resolution via self‐organization neural networks and geometric directional gradient

X‐Ray medical image super‐resolution via self‐organization neural networks and geometric directional gradient

Enhancing Upscaled Image Resolution Using Hybrid Generative Adversarial Network-Enabled Frameworks

Single Image Super Resolution Using Deep Residual Learning

Contact Info

Product

Resources

About