LRT: An Efficient Low-Light Restoration Transformer for Dark Light Field Images

IEEE Transactions on Image Processing, Journal Year: 2023, Volume and Issue: 32, P. 4314 - 4326

Published: Jan. 1, 2023

Light field (LF) images containing information for multiple views have numerous applications, which can be severely affected by low-light imaging. Recent learning-based methods enhancement some disadvantages, such as a lack of noise suppression, complex training process and poor performance in extremely conditions. To tackle these deficiencies while fully utilizing the multi-view information, we propose an efficient Low-light Restoration Transformer (LRT) LF images, with heads to perform intermediate tasks within single network, including denoising, luminance adjustment, refinement detail enhancement, achieving progressive restoration from small scale full scale. Moreover, design angular transformer block view-token scheme model global dependencies, multi-scale spatial encode local each view. address issue insufficient data, formulate synthesis pipeline simulating major sources estimated parameters camera. Experimental results demonstrate that our method achieves state-of-the-art on high efficiency.

Language: Английский

---

Light Field Image Restoration via Latent Diffusion and Multi-View Attention

IEEE Signal Processing Letters, Journal Year: 2024, Volume and Issue: 31, P. 1094 - 1098

Published: Jan. 1, 2024

Light field (LF) images contain information for multiple views. The restoration of degraded LF is great significance various applications. Inspired by the recent achievement denoising diffusion models, we propose a image method based on latent (LD). We design LDUNet with efficient cross-attention modules to integrate features conditional input, and two-stage training strategy, where first trained individual views then fine-tuned injected prior noise. A refinement module jointly in second stage enhance spatial-angular structures. It consists multi-view attention blocks patch-based angular self-attention fuse global view information. Moreover, introduce an enhanced noise loss better prediction auxiliary obtain high-quality images. evaluate our deraining task low-light enhancement task. Our demonstrates superior performance both tasks compared existing methods.

Language: Английский

---

Unsupervised Light Field Depth Estimation via Multi-View Feature Matching With Occlusion Prediction

IEEE Transactions on Circuits and Systems for Video Technology, Journal Year: 2023, Volume and Issue: 34(4), P. 2261 - 2273

Published: Aug. 17, 2023

Depth estimation from light field (LF) images is a fundamental step for numerous applications. Recently, learning-based methods have achieved higher accuracy and efficiency than the traditional methods. However, it costly to obtain sufficient depth labels supervised training. In this paper, we propose an unsupervised framework estimate LF images. First, design disparity network (DispNet) with coarse-to-fine structure predict maps different view combinations. It explicitly performs multi-view feature matching learn correspondences effectively. As occlusions may cause violation of photo-consistency, introduce occlusion prediction (OccNet) maps, which are used as element-wise weights photometric loss solve issue assist learning. With estimated by multiple input combinations, then fusion strategy based on errors effective handling final map accuracy. Experimental results demonstrate that our method achieves superior performance both dense sparse images, also shows better robustness generalization real-world compared other

Language: Английский

---

Semi-Supervised Semantic Segmentation for Light Field Images Using Disparity Information

IEEE Transactions on Image Processing, Journal Year: 2024, Volume and Issue: 33, P. 4516 - 4528

Published: Jan. 1, 2024

Light field (LF) images enable numerous applications due to their ability capture information for multiple views. Semantic segmentation is an essential task LF scene understanding. However, existing supervised methods heavily rely on a large number of pixel-wise annotations. To relieve this problem, we propose semi-supervised semantic method that requires only small subset labeled data and harnesses the disparity information. First, design unsupervised estimation network, which can determine map every view. With estimated maps, generate pseudo-labels along with weight maps peripheral views when labels central are available. We then merge predictions from obtain more reliable unlabeled data, introduce disparity-semantics consistency loss enforce structure similarity. Moreover, develop comprehensive contrastive learning scheme includes pixel-level strategy enhance feature representations object-level improve individual objects. Our demonstrates state-of-the-art performance benchmark dataset under variety training settings achieves comparable trained 1/2 protocol.

Language: Английский

---

LFIENet: Light Field Image Enhancement Network by Fusing Exposures of LF-DSLR Image Pairs

IEEE Transactions on Computational Imaging, Journal Year: 2023, Volume and Issue: 9, P. 620 - 635

Published: Jan. 1, 2023

Plenoptic cameras can record both spatial and angular information of incident rays as 4D light field (LF) images, which have unique advantages in a wide range computer vision graphics applications. However, plenoptic usually suffer from image quality degradation due to limited resolution, very small sub-apertures for sub-views, improper exposure color quantization sensors. Raw macro-pixel LF images captured by are decomposed into an array during decomposition correction would further damage the images. Therefore, sub-views always tricky problems low dynamic range, brightness reduction, deviation missing textural details areas sub-aperture each sub-view. We observed that it is hard tell (tone) ranges DSLR (Digital Single Lens Reflex) Camera better than even same real-world scenes. Thus, instead directly taking accompanying ground truths enhancing we propose unsupervised neural network, called LFIENet, properly fusing exposures LF-DSLR pairs. With help corresponding enhanced contain much abundant extended contrast. Since histogram equalization enhancement able extend improve contrast, Histogram Equalization Attention Module (HEAM) discover over/under-exposed In addition, learning proposed pair dataset. Extensive experiments on various challenging demonstrate effectiveness our network.

Language: Английский

---

CSDM: A Cross-Scale Decomposition Method for Low-Light Image Enhancement

Signal Processing, Journal Year: 2022, Volume and Issue: 202, P. 108752 - 108752

Published: Sept. 6, 2022

Language: Английский

---

PMSNet: Parallel Multi-Scale Network for Accurate Low-Light Light-Field Image Enhancement

IEEE Transactions on Multimedia, Journal Year: 2023, Volume and Issue: 26, P. 2041 - 2055

Published: July 3, 2023

Current low-light light-field (LF) image enhancement algorithms tend to produce blurry results, for (1) loss of spatial details during and (2) inefficient exploitation angular correlations, which helps recover details. Therefore, in this paper, we propose a parallel multi-scale network (PMSNet), attempts process features different scales aggregate the contributions at each layer, thus fully preserve details, integrate multi-resolution 3D convolution streams efficiently utilize correlations. Specifically, PMSNet consists three stages: Stage-I employs modules (MSMs) generate local understanding with aid adjacent views. Notably, MSM retains high-resolution feature extraction minimize Stage-II processes all views encode global information. Based on above extracted information, Stage-III utilizes (3D-MSMs) exploit To validate our idea, comprehensively evaluate performance publicly available datasets. Experimental results show that method is superior current state-of-the-art methods, achieving an average PSNR 24.76 dB.

Language: Английский

---

Unsupervised Disparity Estimation for Light Field Videos

ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Journal Year: 2024, Volume and Issue: unknown

Published: March 18, 2024

Light field (LF) videos contain not only the spatial-angular information but also temporal information, which are useful for disparity estimation. The existing work on estimation LF relies supervised training with labels. To overcome this reliance, we develop an unsupervised framework videos, consists of a matching branch to perform feature and refinement refine maps. Our includes cross-feature fusion module self-attention cross-attention fuse multi-frame features, cost aggregation transformer cross-depth blocks explore global depth dependencies. Moreover, propose left-right consistency strategy estimate occlusion regions input views introduce occlusion-aware photometric loss solve issue. Experimental results demonstrate that our method achieves superior performance compared methods.

Language: Английский

---

Non-Uniform Low-Light Face Image Enhancement Based on Dark Channel Prior and Image Uniform Posterior

IEEE Access, Journal Year: 2024, Volume and Issue: 12, P. 85724 - 85734

Published: Jan. 1, 2024

Language: Английский

---

Combining attention mechanism and Retinex model to enhance low-light images

Computers & Graphics, Journal Year: 2022, Volume and Issue: 104, P. 95 - 105

Published: April 14, 2022

Language: Английский

---