A5046381970- Advanced Image Processing Techniques
- Image and Signal Denoising Methods
- Image Processing Techniques and Applications
- Image Enhancement Techniques
- Advanced Vision and Imaging
- Advanced Image and Video Retrieval Techniques
- Photoacoustic and Ultrasonic Imaging
- Color Science and Applications
- Facial Rejuvenation and Surgery Techniques
- Face recognition and analysis
- Advanced Neural Network Applications
- Advanced Image Fusion Techniques
- Medical Image Segmentation Techniques
- Visual Attention and Saliency Detection
- Image Retrieval and Classification Techniques
- Generative Adversarial Networks and Image Synthesis
- Sparse and Compressive Sensing Techniques
- Facial Nerve Paralysis Treatment and Research
- Image and Video Quality Assessment
- Digital Media Forensic Detection
Alibaba Group (China)
2022-2024
Alibaba Group (United States)
2022-2023
ETH Zurich
2022
University of Bern
2015-2018
Agency for Science, Technology and Research
2014
Pohang University of Science and Technology
2013-2014
Bioinformatics Institute
2014
We present a method to extract video sequence from single motion-blurred image. Motion-blurred images are the result of an averaging process, where instant frames accumulated over time during exposure sensor. Unfortunately, reversing this process is nontrivial. Firstly, destroys temporal ordering frames. Secondly, recovery frame blind deconvolution task, which highly ill-posed. deep learning scheme that gradually reconstructs by sequentially extracting pairs Our main contribution introduce...
This paper reviews the NTIRE 2022 challenge on efficient single image super-resolution with focus proposed solutions and results. The task of was to super-resolve an input a magnification factor ×4 based pairs low corresponding high resolution images. aim design network for that achieved improvement efficiency measured according several metrics including runtime, parameters, FLOPs, activations, memory consumption while at least maintaining PSNR 29.00dB DIV2K validation set. IMDN is set as...
The 3D Lookup Table (3D LUT) is a highly-efficient tool for real-time image enhancement tasks, which models non-linear color transform by sparsely sampling it into discretized lattice. Previous works have made efforts to learn image-adaptive output values of LUTs flexible but neglect the importance strategy. They adopt sub-optimal uniform point allocation, limiting expressiveness learned since (tri-)linear interpolation between points in LUT might fail model local non-linearities transform....
In this paper, we introduce the task of generating a sharp slow-motion video given low frame rate blurry video. We propose data-driven approach, where training data is captured with high camera and images are simulated through an averaging process. While it possible to train neural network recover frames from their average, there no guarantee temporal smoothness for formed video, as estimated independently. To address requirement system two networks: One, DeblurNet, predict keyframes second,...
In this paper we introduce a natural image prior that directly represents Gaussian-smoothed version of the distribution. We include our in formulation restoration as Bayes estimator also allows us to solve noise-blind problems. show gradient corresponds mean-shift vector on addition, learn field using denoising autoencoders, and use it descent approach perform risk minimization. demonstrate competitive results for deblurring, super-resolution, demosaicing.
In this paper we introduce a natural image prior that directly represents Gaussian-smoothed version of the distribution. We include our in formulation restoration as Bayes estimator also allows us to solve noise-blind problems. show gradient corresponds mean-shift vector on addition, learn field using denoising autoencoders, and use it descent approach perform risk minimization. demonstrate competitive results for deblurring, super-resolution, demosaicing.
We present a novel approach to noise-blind deblurring, the problem of deblurring an image with known blur, but unknown noise level. introduce efficient and robust solution based on Bayesian framework using smooth generalization 0-1 loss. A bound allows calculation very high-dimensional integrals in closed form. It avoids degeneracy Maximum a-Posteriori (MAP) estimates leads effective noise-adaptive scheme. Moreover, we drastically accelerate our algorithm by Majorization Minimization (MM)...
Portrait images and photos containing faces are ubiquitous on the web predominant subject of shared via social media. Especially selfie taken with lightweight smartphone cameras susceptible to camera shake. Despite significant progress in field image deblurring over last decade, performance state-of-the-art methods blurry face is still limited. In this work, we present a novel deep learning architecture that designed be computationally fast exploits very wide receptive return sharp even...
Pictures of objects behind a glass are difficult to interpret and understand due the superposition two real images: reflection layer background layer. Separation these layers is challenging ambiguities in assigning texture patterns average color input image one layers. In this paper, we propose novel method reconstruct given single by explicitly handling reconstruction. Our approach combines ability neural networks build priors on large regions with an model that accounts for brightness...
Image matting refers to the problem of foreground extraction from an image and transparency determination pixels. Although other algorithms have been proposed, most are not sufficiently robust obtain satisfactory results in different regions image, such as smooth regions, nonuniform color distribution isolated regions. This paper proposes a novel algorithm that can extract high-quality mattes image. Our proposed combines propagation color-sampling methods. Unlike previous propagation-based...
We propose a method to remove motion blur in single light field captured with moving plenoptic camera. Since is unknown, we resort blind deconvolution formulation, where one aims identify both the point spread function and latent sharp image. Even absence of motion, images by camera are affected non-trivial combination aliasing defocus, which depends on 3D geometry scene. Therefore, deblurring algorithms designed for standard cameras not directly applicable. Moreover, many state art based...
Deep convolutional neural networks have achieved great progress in image denoising tasks. However, their complicated architectures and heavy computational cost hinder deployments on mobile devices. Some recent efforts designing lightweight focus reducing either FLOPs (floating-point operations) or the number of parameters. these metrics are not directly correlated with on-device latency. In this paper, we identify real bottlenecks that affect CNN-based models' runtime performance devices:...
We present a method to extract video sequence from single motion-blurred image. Motion-blurred images are the result of an averaging process, where instant frames accumulated over time during exposure sensor. Unfortunately, reversing this process is nontrivial. Firstly, destroys temporal ordering frames. Secondly, recovery frame blind deconvolution task, which highly ill-posed. deep learning scheme that gradually reconstructs by sequentially extracting pairs Our main contribution introduce...
The formulation of energy minimization Markov random fields has been extensively utilized to infer pixel labels in cellular image segmentation, where a crucial step is specify the data and discontinuity penalty terms functions. In this paper, we propose forest based approach directly learn respective from data. Empirical experiments indicate that our outperforms state-of-the-art methods.
In this paper we propose a solution to blind deconvolution of scene with two layers (foreground/background). We show that the reconstruction support these from single image conventional camera is not possible. As use light field camera. demonstrate captured Lytro can be successfully deblurred. More specifically, consider case space-varying motion blur, where blur magnitude depends on depth changes in scene. Our method employs layered model handles occlusions and partial transparencies due...
Image matting is the extraction of a foreground object from an image and determination transparency each pixel. Matting inherently ill-posed underconstrained problem. Therefore, some assumptions need to be made solve it. Recent methods that provide closed-form solution this problem are based on assumption either local smoothness or nonlocal principle, but they cannot always produce satisfactory results. In paper, we propose K-nearest neighbors (KNN)-based color line model combines preserves...
In this paper, we propose a new matting algorithm using local and nonlocal neighbors. We assume that K nearest neighbors satisfy the color line model RGB distribution of is roughly linear combine assumption with linear. Our assumptions are appropriate for various regions such as those smooth, contain holes or have complex color. Experimental results show proposed method outperforms previous propagation-based methods. Further, it competitive sampling-based methods require sampling learning
Contemporary makeup approaches primarily hinge on unpaired learning paradigms, yet they grapple with the challenges of inaccurate supervision (e.g., face misalignment) and sophisticated facial prompts (including parsing, landmark detection). These prohibit low-cost deployment models, especially mobile devices. To solve above problems, we propose a brand-new paradigm, termed "Data Amplify Learning (DAL)," alongside compact model named "TinyBeauty." The core idea DAL lies in employing...
The 3D Lookup Table (3D LUT) is a highly-efficient tool for real-time image enhancement tasks, which models non-linear color transform by sparsely sampling it into discretized lattice. Previous works have made efforts to learn image-adaptive output values of LUTs flexible but neglect the importance strategy. They adopt sub-optimal uniform point allocation, limiting expressiveness learned since (tri-)linear interpolation between points in LUT might fail model local non-linearities transform....
Image-adaptive lookup tables (LUTs) have achieved great success in real-time image enhancement tasks due to their high efficiency for modeling color transforms. However, they embed the complete transform, including component-independent and component-correlated parts, into only a single type of LUTs, either 1D or 3D, coupled manner. This scheme raises dilemma improving model expressiveness two factors. On one hand, LUTs provide computational but lack critical capability components...
This paper reviews the NTIRE 2022 challenge on efficient single image super-resolution with focus proposed solutions and results. The task of was to super-resolve an input a magnification factor $\times$4 based pairs low corresponding high resolution images. aim design network for that achieved improvement efficiency measured according several metrics including runtime, parameters, FLOPs, activations, memory consumption while at least maintaining PSNR 29.00dB DIV2K validation set. IMDN is...