Youdong Ding’s research while affiliated with Shanghai University and other places

... Liu et al. [8] introduced RAUNA; the structure comprises a decomposition network (DecNet) influenced by algorithmic unrolling and adjustment networks that take into account both global and local brightness. Li et al. [44] introduced UDVN, an enhancement algorithm that effectively handles images with uneven low-light conditions. This algorithm concentrates on the light and shadow details within the image for low-light enhancement tasks, achieving competitive performance in enhancing low-light images. ...

Reference:

Multi-stage residual network with two fold attention mechanisms for low-light image enhancement

... The existing VFI methods can be mainly classified into two categories: flowbased [5,8,17,18] and kernel-based [19][20][21][22]. ...

Reference:

Parallel Spatio-Temporal Attention Transformer for Video Frame Interpolation

... Lu Zou et al. [83] conceptualize 2D poses as graphs, redefining 3D estimation as a graph regression problem, where GCNs infer latent structural relationships within the human body. Bing Yu et al. [84] develop a Perceptual U-shaped Graph Convolutional Network (M-UGCN) using a Ushaped network with map-aware local enhancement, extending the receptive field and intensifying local node interactions across multiple scales to improve 2D-to-3D estimation. Building on this, Hua et al. [85] combine 2D pose estimates from dual views with triangulation to produce an initial 3D pose, subsequently refining it through a Cross-view U-shape Graph Convolutional Network (CV-UGCN) under weak supervision, applicable to any preceding 2D method. ...

Reference:

An In-Depth Analysis of 2D and 3D Pose Estimation Techniques in Deep Learning: Methodologies and Advances

... Although SSL pretext tasks can be designed and employed for many different types of data (e.g., timeseries [16], text [17], video [18] [19], audio [20], point clouds [21], or even multimodal data [22] [23]), this article focuses on image analysis for computer vision applications. Moreover, it focuses on generic SSL methods and not ones explicitly designed for specific tasks (e.g., for multi-view clustering [24], product attribute recognition [25], etc.). The remainder of this paper is organized as follows: Section 2 briefly presents the most common categories of pretext tasks for visual SSL. ...

Reference:

Self-supervised visual learning in the low-data regime: A comparative evaluation

... As in several real-life situations, the ground truth version for the video signals in the new domain is not available, the use of the unsupervised learning techniques for carrying out the task of video-to-video translation is inevitable. Deep neural networks provide the state-of-the-art performances in various computer vision tasks [3][4][5][6][7][8][9], in view of their endto-end learning capability between the input and output data domains. In view of these explanations, the use of the deep unsupervised learning-based schemes seem to be a legitimate choice for designing a high-performance video-to-video translation system. ...

Reference:

FUVT: a deep few-shot unsupervised learning-based video-to-video translation scheme using Kalman filtering and relativistic GAN

... • SGRNet [185] is a two-stage network, which first employs a generator to create a shadow mask by merging foreground and background, and then predicts shadow parameters and fills the shadow area, producing an image with realistic shadows. • Liu et al. [186] enhances shadow generation in image compositing with multi-scale feature enhancement and multi-level feature fusion. This approach improves mask prediction accuracy and minimizes information loss in shadow parameter prediction, leading to enhanced shadow shapes and ranges. ...

Reference:

Unveiling Deep Shadows: A Survey on Image and Video Shadow Detection, Removal, and Generation in the Era of Deep Learning

... The results show that the proposed model produces sharper results closer to the ground truth, with fewer blurring effects and artifacts. Finally, Li et al. [208] (SRAGAN) design a complex GAN with local and global channel and spatial attention modules both in the generator and the discriminator network to capture short-as well as long-range dependencies between pixels. Several experiments proved the superiority of the proposed model, especially at higher scaling factors. ...

Reference:

Deep Learning for Downscaling Remote Sensing Images: Fusion and Super-Resolution

... Chu et al. [11] proposed a two-stage network that first extracts the subtitle mask using a mask extraction network, then feeds the predicted mask and video frame into the generator to remove subtitles. Tu et al. [33] propose a lightweight mask extraction network that uses gated convolutions to generate unsubtitled videos. Although BVDNet [10] has small model parameters, it does not perform well in removing subtitles. ...

Reference:

Deep video decaptioning based on multi-scale large kernel attention

... Despite data limitations, recent research has explored different SLLIE [17] approaches. The end-to-end learningbased methods [2,9,11,15,22,26,29,30,39,43,53] aim to directly map low-light images to well-lit counterparts, often integrating Convolutional Neural Networks (CNNs). Several recent methods leverage transformers [7,41,52] to enhance the receptive field of vanilla convolutions for learning spatial information, inspired by advances in image restoration [48,57,58]. ...

Reference:

Illuminating Darkness: Enhancing Real-world Low-light Scenes with Smartphone Images

... Assuming x is a natural image, φ is a histogram matching operator that can be be widely applied to image color transfer 49,50 , color correction 51,52 , and contrast enhancement 53 . For an input image x = (x1, x2, . . . ...

Reference:

An alternating multiple residual Wasserstein regularization model for Gaussian image denoising