Deep Residual Learning for Image Recognition: A Survey

Applied Sciences, Journal Year: 2022, Volume and Issue: 12(18), P. 8972 - 8972

Published: Sept. 7, 2022

Deep Residual Networks have recently been shown to significantly improve the performance of neural networks trained on ImageNet, with results beating all previous methods this dataset by large margins in image classification task. However, meaning these impressive numbers and their implications for future research are not fully understood yet. In survey, we will try explain what are, how they achieve excellent results, why successful implementation practice represents a significant advance over existing techniques. We also discuss some open questions related residual learning as well possible applications beyond ImageNet. Finally, issues that still need be resolved before deep can applied more complex problems.

Language: Английский

---

Remote sensing image super-resolution and object detection: Benchmark and state of the art

Expert Systems with Applications, Journal Year: 2022, Volume and Issue: 197, P. 116793 - 116793

Published: March 1, 2022

Language: Английский

---

Image super-resolution: A comprehensive review, recent trends, challenges and applications

Information Fusion, Journal Year: 2022, Volume and Issue: 91, P. 230 - 260

Published: Oct. 14, 2022

Language: Английский

---

High-fidelity reconstruction of turbulent flow from spatially limited data using enhanced super-resolution generative adversarial network

Physics of Fluids, Journal Year: 2021, Volume and Issue: 33(12)

Published: Dec. 1, 2021

In this study, a deep learning-based approach is applied with the aim of reconstructing high-resolution turbulent flow fields using minimal data. A multi-scale enhanced super-resolution generative adversarial network physics-based loss function introduced as model to reconstruct fields. The capability laminar flows examined data around square cylinder. results reveal that can accurately reproduce even when limited spatial information provided. case channel used assess ability wall-bounded instantaneous and statistical obtained from agree well ground truth data, indicating successfully learn map coarse once. Furthermore, computational cost proposed model, which carefully, found be effectively low. This demonstrates high-fidelity training physics-guided network-based models practically efficient in extremely

Language: Английский

---

Small Object Detection in Remote Sensing Images with Residual Feature Aggregation-Based Super-Resolution and Object Detector Network

Remote Sensing, Journal Year: 2021, Volume and Issue: 13(9), P. 1854 - 1854

Published: May 10, 2021

This paper deals with detecting small objects in remote sensing images from satellites or any aerial vehicle by utilizing the concept of image super-resolution for resolution enhancement using a deep-learning-based detection method. provides rationale improving current (SR) framework incorporating cyclic generative adversarial network (GAN) and residual feature aggregation (RFA) to improve performance. The novelty method is threefold: first, proposed, independent final object detector used research, i.e., YOLOv3 could be replaced Faster R-CNN perform detection; second, was generator, which significantly improved performance as RFA detected complex features; third, whole transformed into GAN. GAN YOLO termed SRCGAN-RFA-YOLO, compared accuracies other methods. Rigorous experiments on both satellite (ISPRS Potsdam, VAID, Draper Satellite Image Chronology datasets) were performed, results showed that increased methods spatial enhancement; an IoU 0.10, AP 0.7867 achieved scale factor 16.

Language: Английский

---

TranSMS: Transformers for Super-Resolution Calibration in Magnetic Particle Imaging

IEEE Transactions on Medical Imaging, Journal Year: 2022, Volume and Issue: 41(12), P. 3562 - 3574

Published: July 11, 2022

Magnetic particle imaging (MPI) offers exceptional contrast for magnetic nanoparticles (MNP) at high spatio-temporal resolution. A common procedure in MPI starts with a calibration scan to measure the system matrix (SM), which is then used set up an inverse problem reconstruct images of MNP distribution during subsequent scans. This enables reconstruction sensitively account various imperfections. Yet time-consuming SM measurements have be repeated under notable changes properties. Here, we introduce novel deep learning approach accelerated based on Transformers super-resolution (TranSMS). Low-resolution are performed using large samples improved signal-to-noise ratio efficiency, and high-resolution super-resolved via model-based learning. TranSMS leverages vision transformer module capture contextual relationships low-resolution input images, dense convolutional localizing image features, data-consistency ensure measurement fidelity. Demonstrations simulated experimental data indicate that significantly improves recovery 64-fold acceleration two-dimensional imaging.

Language: Английский

---

Three-dimensional ESRGAN for super-resolution reconstruction of turbulent flows with tricubic interpolation-based transfer learning

Physics of Fluids, Journal Year: 2022, Volume and Issue: 34(12)

Published: Nov. 23, 2022

Turbulence is a complicated phenomenon because of its chaotic behavior with multiple spatiotemporal scales. also has irregularity and diffusivity, making predicting reconstructing turbulence more challenging. This study proposes deep-learning approach to reconstruct three-dimensional (3D) high-resolution turbulent flows from spatially limited data using 3D enhanced super-resolution generative adversarial networks (3D-ESRGAN). In addition, novel transfer-learning method based on tricubic interpolation employed. Turbulent channel flow at friction Reynolds numbers Reτ = 180 500 were generated by direct numerical simulation (DNS) used estimate the performance model as well that interpolation-based transfer learning. The results, including instantaneous velocity fields statistics, show reconstructed agree reference DNS data. findings indicate proposed 3D-ESRGAN can even training

Language: Английский

---

Super-resolution reconstruction of turbulent flow fields at various Reynolds numbers based on generative adversarial networks

Physics of Fluids, Journal Year: 2022, Volume and Issue: 34(1)

Published: Jan. 1, 2022

This study presents a deep learning-based framework to recover high-resolution turbulent velocity fields from extremely low-resolution data at various Reynolds numbers by utilizing the concept of generative adversarial networks. A multiscale enhanced super-resolution network is applied as model reconstruct fields, and direct numerical simulation channel flow with large longitudinal ribs are used evaluate performance model. The found have capacity accurately two different down-sampling factors in terms instantaneous two-point correlations, turbulence statistics. results further reveal that able fall within range training numbers.

Language: Английский

---

Reconstructing high fidelity digital rock images using deep convolutional neural networks

Scientific Reports, Journal Year: 2022, Volume and Issue: 12(1)

Published: March 11, 2022

Abstract Imaging methods have broad applications in geosciences. Scanning electron microscopy (SEM) and micro-CT scanning been applied for studying various geological problems. Despite significant advances imaging capabilities, image processing algorithms, acquiring high-quality data from images is still challenging time-consuming. Obtaining a 3D representative volume tight rock sample takes days to weeks. Image artifacts such as noise further complicate the use of determination properties. In this study, we present several convolutional neural networks (CNN) rapid denoising, deblurring super-resolving digital images. Such an approach enables larger samples, which turn improves statistical relevance subsequent analysis. We demonstrate application CNNs restoration applicable scientific imaging. The results show that can be denoised without priori knowledge with great confidence. Furthermore, how attaching end-to-end fashion improve final quality reconstruction. Our experiments SEM CT scan types super-resolution performed simultaneously.

Language: Английский

---

Hitchhiker's Guide to Super-Resolution: Introduction and Recent Advances

IEEE Transactions on Pattern Analysis and Machine Intelligence, Journal Year: 2023, Volume and Issue: 45(8), P. 9862 - 9882

Published: Feb. 10, 2023

With the advent of Deep Learning (DL), Super-Resolution (SR) has also become a thriving research area. However, despite promising results, field still faces challenges that require further e.g., allowing flexible upsampling, more effective loss functions, and better evaluation metrics. We review domain SR in light recent advances, examine state-of-the-art models such as diffusion (DDPM) transformer-based models. present critical discussion on contemporary strategies used SR, identify yet unexplored directions. complement previous surveys by incorporating latest developments uncertainty-driven losses, wavelet networks, neural architecture search, novel normalization methods, latests techniques. include several visualizations for methods throughout each chapter order to facilitate global understanding trends field. This is ultimately aimed at helping researchers push boundaries DL applied SR.

Language: Английский

---