SRLibrary: Comparing different loss functions for super-resolution over various convolutional architectures

Anagün, Yıldıray; Işık, Şahin; Seke, Erol

doi:10.1016/j.jvcir.2019.03.027

Cited by 57 publications

(11 citation statements)

References 11 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The reason of why we used L1 loss, is explained in the thesis (Anagün 2018). In this study, the L1 loss function is faster than the other loss functions and gives effective results for solving the SR problem (Anagun et al 2019).…”

Section: Methods and Toolsmentioning

confidence: 95%

Contribution Analysis of Optimization Methods on Super-Resolution

Anagün¹,

Işık²

2021

Afyon Kocatepe University Journal of Sciences and Engineering

View full text Add to dashboard Cite

In this study, the benefits of choosing a robust optimization function with super resolution are analyzed. For this purpose, the different optimizers are included in the simple Convolutional Neural Network (CNN) architecture SRNET, to reveal the performance of the each method. Findings of this research provides that Adam and Nadam optimizers are robust when compared to (Stochastic Gradient Descent) SGD, Adagrad, Adamax and RMSprop. After experimental simulations, we have achieved the 35.91 (dB)/0.9960 and 35.97 (dB)/0.9961 accuracy rates on Set5 images from Adam and Nadam optimizers, respectively (9-1-5 network structure and filter sizes 128 and 64). These results show that selected optimization function for the CNN model plays an important role in increasing the accuracy rate in the super-resolution problem.

show abstract

Section: Methods and Toolsmentioning

confidence: 95%

Contribution Analysis of Optimization Methods on Super-Resolution

Anagün¹,

Işık²

2021

Afyon Kocatepe University Journal of Sciences and Engineering

View full text Add to dashboard Cite

show abstract

“…Finally, we apply the pixel shuffling operation to increase the spatial resolution of LF features, and further employ a 3×3 convolution to obtain the super-resolved LF image L SR . Following most existing works [60,59,35,61,56,37,77,78,71], we use the L 1 loss function to train our network due to its robustness to outliers [2].…”

Section: Feature Upsamplingmentioning

confidence: 99%

Learning Non-Local Spatial-Angular Correlation for Light Field Image Super-Resolution

Liang¹,

Wang²,

Wang³

et al. 2023

Preprint

View full text Add to dashboard Cite

Exploiting spatial-angular correlation is crucial to light field (LF) image super-resolution (SR), but is highly challenging due to its non-local property caused by the disparities among LF images. Although many deep neural networks (DNNs) have been developed for LF image SR and achieved continuously improved performance, existing methods cannot well leverage the long-range spatialangular correlation and thus suffer a significant performance drop when handling scenes with large disparity variations. In this paper, we propose a simple yet effective method to learn the non-local spatial-angular correlation for LF image SR. In our method, we adopt the epipolar plane image (EPI) representation to project the 4D spatialangular correlation onto multiple 2D EPI planes, and then develop a Transformer network with repetitive self-attention operations to learn the spatial-angular correlation by modeling the dependencies between each pair of EPI pixels. Our method can fully incorporate the information from all angular views while achieving a global receptive field along the epipolar line. We conduct extensive experiments with insightful visualizations to validate the effectiveness of our method. Comparative results on five public datasets show that our method not only achieves state-of-the-art SR performance, but also performs robust to disparity variations. Code is publicly available at https://github.com/ ZhengyuLiang24/EPIT.

show abstract

“…We use Charbonnier loss [3] to quantify the error between the high-resolution output and the given ground truth image. Charbonnier loss is known to be insensitive to outliers and for super-resolution tasks, experimental evaluation has shown that it provides better PSNR/SSIM accuracies over other conventional loss functions [1].…”

Section: Loss Functionmentioning

confidence: 99%