A5031957432- Advanced Vision and Imaging
- 3D Shape Modeling and Analysis
- Computer Graphics and Visualization Techniques
- Advanced Image Processing Techniques
- Image Enhancement Techniques
- Advanced Image Fusion Techniques
- 3D Surveying and Cultural Heritage
- Remote-Sensing Image Classification
- Human Pose and Action Recognition
- Video Coding and Compression Technologies
- Advanced Neural Network Applications
- Image and Video Quality Assessment
- Optical measurement and interference techniques
- Face and Expression Recognition
- Video Analysis and Summarization
- Image and Signal Denoising Methods
- Video Surveillance and Tracking Methods
- Advanced Image and Video Retrieval Techniques
- Sparse and Compressive Sensing Techniques
- Visual Attention and Saliency Detection
- Advanced Data Compression Techniques
- Remote Sensing and Land Use
- Remote Sensing and LiDAR Applications
- Advanced Numerical Analysis Techniques
- Image Processing and 3D Reconstruction
City University of Hong Kong
2017-2025
Xi'an University of Technology
2024
City University of Hong Kong, Shenzhen Research Institute
2017-2023
Griffith University
2023
Qingdao University
2018-2022
Affiliated Hospital of Qingdao University
2022
Jiangsu University of Science and Technology
2022
South China University of Technology
2021
Southeast University
2021
Nanyang Technological University
2012-2020
The paper presents a novel method, Zero-Reference Deep Curve Estimation (Zero-DCE), which formulates light enhancement as task of image-specific curve estimation with deep network. Our method trains lightweight network, DCE-Net, to estimate pixel-wise and high-order curves for dynamic range adjustment given image. is specially designed, considering pixel value range, monotonicity, differentiability. Zero-DCE appealing in its relaxed assumption on reference images, i.e., it does not require...
Underwater image enhancement has been attracting much attention due to its significance in marine engineering and aquatic robotics. Numerous underwater algorithms have proposed the last few years. However, these are mainly evaluated using either synthetic datasets or selected real-world images. It is thus unclear how would perform on images acquired wild we could gauge progress field. To bridge this gap, present first comprehensive perceptual study analysis of large-scale In paper, construct...
Underwater images suffer from color casts and low contrast due to wavelength- distance-dependent attenuation scattering. To solve these two degradation issues, we present an underwater image enhancement network via medium transmission-guided multi-color space embedding, called Ucolor. Concretely, first propose a encoder network, which enriches the diversity of feature representations by incorporating characteristics different spaces into unified structure. Coupled with attention mechanism,...
In rate-distortion optimization, the encoder settings are determined by maximizing a reconstruction quality measure subject to constraint on bit rate. One of main challenges this approach is define that can be computed with low computational cost and which correlates well perceptual quality. While several measures fulfil these two criteria have been developed for images video, no such one exists 3D point clouds. We address limitation video-based cloud compression (V-PCC) standard proposing...
We propose a generative adversarial network for point cloud upsampling, which can not only make the upsampled points evenly distributed on underlying surface but also efficiently generate clean high frequency regions. The generator of our includes dynamic graph hierarchical residual aggregation unit and feature extraction respectively. former extracts multiscale point-wise descriptive features, while latter captures rich details with residuals. To neat edges, discriminator uses filter to...
Convolutional neural network (CNN) is well known for its capability of feature learning and has made revolutionary achievements in many applications, such as scene recognition target detection. In this paper, hyperspectral images explored by constructing a five-layer CNN classification (C-CNN). The proposed C-CNN constructed including recent advances deep area, batch normalization, dropout, parametric rectified linear unit (PReLU) activation function. addition, both spatial context spectral...
The elderly population is increasing rapidly all over the world. One major risk for people fall accidents, especially those living alone. In this paper, we propose a robust detection approach by analyzing tracked key joints of human body using single depth camera. Compared to rivals that rely on RGB inputs, proposed scheme independent illumination lights and can work even in dark room. our scheme, pose-invariant randomized decision tree algorithm joint extraction, which requires low...
Salient object detection from RGB-D images is an important yet challenging vision task, which aims at detecting the most distinctive objects in a scene by combining color information and depth constraints. Unlike prior fusion manners, we propose attention steered interweave network (ASIF-Net) to detect salient objects, progressively integrates cross-modal cross-level complementarity RGB image corresponding map via steering of mechanism. Specifically, complementary features are jointly...
Rain removal is important for improving the robustness of outdoor vision based systems. Current rain methods show limitations either complex dynamic scenes shot from fast moving cameras, or under torrential fall with opaque occlusions. We propose a novel derain algorithm, which applies superpixel (SP) segmentation to decompose scene into depth consistent units. Alignment contents are done at SP level, proves be robust towards occlusion and camera motion. Two alignment output tensors, i.e.,...
Light field (LF) photography is an emerging paradigm for capturing more immersive representations of the real-world. However, arising from inherent trade-off between angular and spatial dimensions, resolution LF images captured by commercial micro-lens based cameras are significantly constrained. In this paper, we propose effective efficient end-to-end convolutional neural network models spatially super-resolving images. Specifically, proposed have hourglass shape, which allows feature...
Depth information has been demonstrated to be useful for saliency detection. However, the existing methods RGBD detection mainly focus on designing straightforward and comprehensive models, while ignoring transferable ability of RGB models. In this article, we propose a novel depth-guided transformation model (DTM) going from saliency. The proposed includes three components, that is: 1) multilevel initialization; 2) refinement; 3) optimization with depth constraints. explicit feature is...
This paper reviews the second challenge on spectral reconstruction from RGB images, i.e., recovery of whole- scene hyperspectral (HS) information a 3-channel image. As in previous challenge, two tracks were provided: (i) "Clean" track where HS images are estimated noise-free RGBs, themselves calculated numerically using ground-truth and supplied sensitivity functions (ii) "Real World" track, simulating capture by an uncalibrated unknown camera, recovered noisy JPEG-compressed images. A new,...
Light field (LF) images acquired by hand-held devices usually suffer from low spatial resolution as the limited sampling resources have to be shared with angular dimension. LF super-resolution (SR) thus becomes an indispensable part of camera processing pipeline. The high-dimensionality characteristic and complex geometrical structure makes problem more challenging than traditional single-image SR. performance existing methods are still they fail thoroughly explore coherence among views...
In recent years, point clouds have become increasingly popular for representing three-dimensional (3D) visual objects and scenes. To efficiently store transmit clouds, compression methods been developed, but they often result in a degradation of quality. reduce color distortion we propose graph-based quality enhancement network (GQE-Net) that uses geometry information as an auxiliary input graph convolution blocks to extract local features efficiently. Specifically, use parallel-serial...
Depth estimation is a fundamental problem for light field photography applications. Numerous methods have been proposed in recent years, which either focus on crafting cost terms more robust matching, or analyzing the geometry of scene structures embedded epipolar-plane images. Significant improvements made overall depth error; however, current state-of-the-art still show limitations handling intricate occluding and complex scenes with multiple occlusions. To address these challenging...
Point cloud based 3D visual representation is becoming popular due to its ability exhibit the real world in a more comprehensive and immersive way. However, under limited network bandwidth, it very challenging communicate this kind of media huge data volume. Therefore, MPEG have launched standardization for point compression (PCC), proposed three model categories, i.e., TMC1, TMC2, TMC3. Because geometry methods TMC1 TMC3 are similar, further merged into new platform namely TMC13. In paper,...
In this paper, an accurate and efficient full-reference image quality assessment (IQA) model using the extracted Gabor features, called feature-based (GFM), is proposed for conducting objective evaluation of screen content images (SCIs). It well-known that filters are highly consistent with response human visual system (HVS), HVS sensitive to edge information. Based on these facts, imaginary part filter has odd symmetry yields detection exploited luminance reference distorted SCI extracting...
The acquisition of light field images with high angular resolution is costly. Although many methods have been proposed to improve the a sparsely-sampled field, they always focus on small baseline, which captured by consumer camera. By making full use intrinsic geometry information fields, in this paper we propose an end-to-end learning-based approach aiming at angularly super-resolving large baseline. Our model consists two learnable modules and physically-based module. Specifically, it...
In this article, we propose a novel self-training approach named Crowd-SDNet that enables typical object detector trained only with point-level annotations (i.e., objects are labeled points) to estimate both the center points and sizes of crowded objects. Specifically, during training, utilize available point supervise estimation directly. Based on locally-uniform distribution assumption, initialize pseudo from supervisory information, which then leveraged guide regression via...
Recently, 3D point cloud is becoming popular due to its capability represent the real world for advanced content modality in modern communication systems. In view of wide applications, especially immersive towards human perception, quality metrics clouds are essential. Existing evaluations rely on a full or certain portion original cloud, which severely limits their applications. To overcome this problem, we propose novel deep learning-based no reference assessment method, namely PQA-Net....
Recent imaging technologies are rapidly evolving for sampling richer and more immersive representations of the 3D world. And one emerging light field (LF) cameras based on micro-lens arrays. To record directional information rays, a much larger storage space transmission bandwidth required by LF image as compared with conventional 2D similar spatial dimension, compression data becomes vital part its application. In this paper, we propose codec that fully exploits intrinsic geometry between...
Compared with conventional color images, light field images (LFIs) contain richer scene information, which allows a wide range of interesting applications. However, such additional information is obtained at the cost generating substantially more data, poses challenges to both data storage and transmission. In this paper, we propose new hybrid framework for effective compression LFIs. The proposed takes particular characteristics LFIs into account so that inter- intra-view correlations can...
Arising from the various object types and scales, diverse imaging orientations, cluttered backgrounds in optical remote sensing image (RSI), it is difficult to directly extend success of salient detection for nature scene RSI. In this paper, we propose an end-to-end deep network called LV-Net based on shape architecture, which detects objects RSIs a purely data-driven fashion. The proposed consists two key modules, i.e., two-stream pyramid module (L-shaped module) encoder-decoder with nested...