Light Field Image Super-Resolution via Mutual Attention Guidance DOI Creative Commons
Zijian Wang, Yao Lu

IEEE Access, Journal Year: 2021, Volume and Issue: 9, P. 129022 - 129031

Published: Jan. 1, 2021

Deep learning-based methods have prompted light field image super-resolution to achieve significant progress. However, most of them ignore aligning different sub-aperture features before aggregation, resulting in sub-optimal results. We aim propose an efficient feature alignment method for aggregation. To this end, we develop a mutual attention mechanism and guidance block (MAG). MAG achieves the between center surrounding with module (CAG) (SAG). CAG aligns center-view surrounding-view generates refined feature, while SAG original implement bidirectional center-view, view alignment. Based on MAG, build Light Field Mutual Attention Guidance Network (LF-MAGNet) constructed by multiple MAGs cascade manner. Experiments are performed commonly-used benchmarks. Our superior qualitative quantitative results other state-of-the-art methods, which demonstrate effectiveness our LF-MAGNet.

Language: Английский

Advanced Optical Imaging Technologies for Microplastics Identification: Progress and Challenges DOI Creative Commons
Yanmin Zhu, Yuxing Li, Jianqing Huang

et al.

Advanced Photonics Research, Journal Year: 2024, Volume and Issue: 5(11)

Published: July 22, 2024

Global concern about microplastic (MP) and nanoplastic (NP) particles is continuously rising with their proliferation worldwide. Effective identification methods for MP NP pollution monitoring are highly needed, but due to different requirements technical challenges, much of the work still in progress. Herein, advanced optical imaging systems that successfully applied or have potential focused on. Compared chemical thermal analyses, unique advantages being nondestructive noncontact allow fast detection without complex sample preprocessing. Furthermore, they capable revealing morphology, anisotropy, material characteristics quick robust detection. This review aims present a comprehensive discussion relevant systems, emphasizing operating principles, strengths, drawbacks. Multiple comparisons analyses among these technologies conducted order provide practical guidelines researchers. In addition, combination other alternative described representative portable devices highlighted. Together, shed light on prospects long‐term environmental protection.

Language: Английский

Citations

7
Robust Reconstruction With Deep Learning to Handle Model Mismatch in Lensless Imaging DOI
Tianjiao Zeng, Edmund Y. Lam

IEEE Transactions on Computational Imaging, Journal Year: 2021, Volume and Issue: 7, P. 1080 - 1092

Published: Jan. 1, 2021

Mask-based lensless imaging is an emerging modality, which replaces the lenses with optical elements and makes use of computation to reconstruct images from multiplexed measurements. Most existing reconstruction algorithms are implemented assuming that forward process a convolution operation, point spread function based on system model. In reality, there model mismatch, leading inferior image results. this paper, we investigate impact mismatch in mask-based for first time, illustrate accumulated artifacts information loss due error state-of-the-art approaches, perform model-based learning-based enhancement separate stages. To overcome this, develop novel physics-informed deep learning architecture aims at addressing such error. The proposed hybrid network consists both unrolled optimization apply physics layers correction. Besides cascaded network, introduce data-driven branch parallel, making input measurement all intermediate outputs correct bias compensate mismatch. effectiveness robustness compensation referred as MMCN, demonstrated real images. Experimental results show noticeably better performance MMCN compared alternative methods.

Language: Английский

Citations

30
Radiograph-comparable image synthesis for spine alignment analysis using deep learning with prospective clinical validation DOI Creative Commons
Nan Meng, Kenneth K. Wong, Moxin Zhao

et al.

EClinicalMedicine, Journal Year: 2023, Volume and Issue: 61, P. 102050 - 102050

Published: June 22, 2023

Adolescent idiopathic scoliosis (AIS) is the most common type of spinal disorder affecting children. Clinical screening and diagnosis require physical radiographic examinations, which are either subjective or increase radiation exposure. We therefore developed validated a radiation-free portable system device utilising light-based depth sensing deep learning technologies to analyse AIS by landmark detection image synthesis.Consecutive patients with attending two local clinics in Hong Kong between October 9, 2019, May 21, 2022, were recruited. Patients excluded if they had psychological and/or systematic neural disorders that could influence compliance study mobility patients. For each participant, Red Green Blue-Depth (RGBD) nude back was collected using our in-house device. Manually labelled landmarks alignment parameters spine surgeons considered as ground truth (GT). Images from training internal validation cohorts (n = 1936) used develop models. The model then prospectively on another cohort 302) same demographic properties cohort. evaluated prediction accuracy well performance radiograph-comparable (RCI) synthesis. obtained RCIs contain sufficient anatomical information can quantify disease severities curve types.Our consistently high predicting less than 4-pixel error regarding mean Euclidian Manhattan distance. synthesized RCI for severity classification achieved sensitivity negative predictive value over 0.909 0.933, 0.974 0.908, specialists' manual assessment results real radiographs GT. estimated Cobb angle strong correlation GT angles (R2 0.984, p < 0.001).The medical powered techniques provide instantaneous harmless analysis has potential integration into routine adolescents.Innovation Technology Fund (MRP/038/20X), Health Services Research (HMRF) 08192266.

Language: Английский

Citations

12
Unsupervised light field disparity estimation using confidence weight and occlusion-aware DOI
Bo Xiao, Xiujing Gao, Huadong Zheng

et al.

Optics and Lasers in Engineering, Journal Year: 2025, Volume and Issue: 189, P. 108928 - 108928

Published: March 5, 2025

Language: Английский

Citations

0
Learning to restore light fields under low-light imaging DOI

Shansi Zhang,

Edmund Y. Lam

Neurocomputing, Journal Year: 2021, Volume and Issue: 456, P. 76 - 87

Published: May 25, 2021

Language: Английский

Citations

22
An effective decomposition-enhancement method to restore light field images captured in the dark DOI

Shansi Zhang,

Edmund Y. Lam

Signal Processing, Journal Year: 2021, Volume and Issue: 189, P. 108279 - 108279

Published: Aug. 4, 2021

Language: Английский

Citations

16
Learning-based light field imaging: an overview DOI Creative Commons
Saeed Mahmoudpour, Carla L. Pagliari, Peter Schelkens

et al.

EURASIP Journal on Image and Video Processing, Journal Year: 2024, Volume and Issue: 2024(1)

Published: May 30, 2024

Abstract Conventional photography can only provide a two-dimensional image of the scene, whereas emerging imaging modalities such as light field enable representation higher dimensional visual information by capturing rays from different directions. Light fields immersive experiences, sense presence in and enhance vision tasks. Hence, research into processing methods has become increasingly popular. It does, however, come at cost data volume computational complexity. With growing deployment machine-learning deep architectures applications, paradigm shift toward learning-based approaches also been observed design methods. Various are developed to process high efficiently for tasks while improving performance. Taking account diversity deployed frameworks, it is necessary survey scattered works domain gain insight current trends challenges. This paper aims review existing solutions summarize most promising frameworks. Moreover, evaluation available datasets highlighted. Lastly, concludes with brief outlook future

Language: Английский

Citations

1
Efficiency–Accuracy Trade-Off in Light Field Estimation with Cost Volume Construction and Aggregation DOI Creative Commons
Bo Xiao, Stuart Perry, Xiujing Gao

et al.

Sensors, Journal Year: 2024, Volume and Issue: 24(11), P. 3583 - 3583

Published: June 1, 2024

The Rich spatial and angular information in light field images enables accurate depth estimation, which is a crucial aspect of environmental perception. However, the abundance also leads to high computational costs memory pressure. Typically, selectively pruning some can significantly improve efficiency but at expense reduced estimation accuracy pruned model, especially low-texture regions occluded areas where diversity reduced. In this study, we propose lightweight disparity model that balances speed enhances textureless regions. We combined cost matching methods based on absolute difference correlation construct volumes, improving both robustness. Additionally, developed multi-scale fusion architecture, employing 3D convolutions UNet-like structure handle different scales. This method effectively integrates across scales, utilizing UNet for efficient completion thus yielding more precise maps. Extensive testing shows our achieves par with most existing methods, yet double accuracy. Moreover, approach comparable current highest-accuracy an order magnitude improvement performance.

Language: Английский

Citations

1
Reivew of Light Field Image Super-Resolution DOI Open Access
Li Yu, Yunpeng Ma,

Song Hong

et al.

Electronics, Journal Year: 2022, Volume and Issue: 11(12), P. 1904 - 1904

Published: June 17, 2022

Currently, light fields play important roles in industry, including 3D mapping, virtual reality and other fields. However, as a kind of high-latitude data, field images are difficult to acquire store. Thus, the study super-resolution is great importance. Compared with traditional 2D planar images, 4D contain information from different angles scene, thus needs be performed not only spatial domain but also angular domain. In early days research, many solutions for image super-resolution, such Gaussian models sparse representations, were used super-resolution. With development deep learning, based on deep-learning techniques becoming increasingly common gradually replacing methods. this paper, current research methods deep-learning-based methods, outlined discussed separately. This paper lists publicly available datasets compares performance various these well analyses importance its future development.

Language: Английский

Citations

5
Fast Light Field Angular Resolution Enhancement Using Convolutional Neural Network DOI Creative Commons
Xingzheng Wang, Senlin You, Yongqiang Zan

et al.

IEEE Access, Journal Year: 2021, Volume and Issue: 9, P. 30216 - 30224

Published: Jan. 1, 2021

Current light field angular super-resolution algorithm based on deep learning has excessive computation cost and low operational efficiency, for sequentially up-sampling each lenslet region of the image. In this paper, we propose a novel convolutional neural network to fastly enhance resolution, via wholesale regions. Firstly, simultaneously extracts information all regions input Then, from extracted information, four feature maps are predicted. Especially, resolution map is same as that Finally, integrate into one image, by referring arrangement in The experimental results verify effectiveness our proposed method. We only need 11.95s enhance(actually double) image with 2562 × 3724 pixels, which surpasses 20 times faster than state-of-the-art Meanwhile, method also achieves average PSNR gains 0.39 dB.

Language: Английский

Citations

6