A5059112338- Image and Signal Denoising Methods
- Advanced Image Processing Techniques
- Photoacoustic and Ultrasonic Imaging
- Advanced MRI Techniques and Applications
- Image Processing Techniques and Applications
- Advanced Image Fusion Techniques
- Sparse and Compressive Sensing Techniques
- Advanced Fluorescence Microscopy Techniques
- Medical Imaging Techniques and Applications
- Generative Adversarial Networks and Image Synthesis
- Nuclear Physics and Applications
- Fire effects on concrete materials
- Advanced Image and Video Retrieval Techniques
- Advanced Vision and Imaging
- Optical Coherence Tomography Applications
- Structural Response to Dynamic Loads
- Innovative concrete reinforcement materials
- Concrete and Cement Materials Research
- Cell Image Analysis Techniques
- Geomechanics and Mining Engineering
- Image Retrieval and Classification Techniques
- Digital Media Forensic Detection
- Advanced Neural Network Applications
- Multimodal Machine Learning Applications
- Remote-Sensing Image Classification
Tsinghua University
2020-2024
University Town of Shenzhen
2021-2024
Xi'an University of Architecture and Technology
2022-2023
Tsinghua–Berkeley Shenzhen Institute
2021-2023
Shenzhen Institute of Information Technology
2021
Hyperspectral image (HSI) reconstruction aims to recover the 3D spatial-spectral signal from a 2D measurement in coded aperture snapshot spectral imaging (CASSI) system. The HSI representations are highly similar and correlated across dimension. Modeling inter-spectra interactions is beneficial for reconstruction. However, existing CNN-based methods show limitations capturing spectral-wise similarity long-range dependencies. Besides, information modulated by (physical mask) CASSI....
The rapid development of deep learning provides a better solution for the end-to-end reconstruction hyperspectral image (HSI). However, existing learning-based methods have two major defects. Firstly, networks with self-attention usually sacrifice internal resolution to balance model performance against complexity, losing fine-grained high-resolution (HR) features. Secondly, even if optimization focusing on spatial-spectral domain (SDL) converges ideal solution, there is still significant...
Abstract A fundamental challenge in fluorescence microscopy is the photon shot noise arising from inevitable stochasticity of detection. Noise increases measurement uncertainty and limits imaging resolution, speed sensitivity. To achieve high-sensitivity beyond shot-noise limit, we present DeepCAD-RT, a self-supervised deep learning method for real-time suppression. Based on our previous framework DeepCAD, reduced number network parameters by 94%, memory consumption 27-fold processing time...
Abstract Fluorescence imaging with high signal-to-noise ratios has become the foundation of accurate visualization and analysis biological phenomena. However, inevitable noise poses a formidable challenge to sensitivity. Here we provide spatial redundancy denoising transformer (SRDTrans) remove from fluorescence images in self-supervised manner. First, sampling strategy based on is proposed extract adjacent orthogonal training pairs, which eliminates dependence speed. Second, designed...
Existing deep learning real denoising methods require a large amount of noisy-clean image pairs for supervision. Nonetheless, capturing dataset is an unacceptable expensive and cumbersome procedure. To alleviate this problem, work investigates how to generate realistic noisy images. Firstly, we formulate simple yet reasonable noise model that treats each pixel as random variable. This splits the generation problem into two sub-problems: domain alignment alignment. Subsequently, propose novel...
High-quality magnetic resonance (MR) images afford more detailed information for reliable diagnoses and quantitative image analyses. Given low-resolution (LR) images, the deep convolutional neural network (CNN) has shown its promising ability super-resolution (SR). The LR MR usually share some visual characteristics: structural textures of different sizes, edges with high correlation, less informative background. However, multi-scale features are reconstruction, while background is smooth....
Exploiting similar and sharper scene patches in spatio-temporal neighborhoods is critical for video deblurring. However, CNN-based methods show limitations capturing long-range dependencies modeling non-local self-similarity. In this paper, we propose a novel framework, Flow-Guided Sparse Transformer (FGST), FGST, customize self-attention module, Window-based Multi-head Self-Attention (FGSW-MSA). For each $query$ element on the blurry reference frame, FGSW-MSA enjoys guidance of estimated...
Many studies have concentrated on constructing supervised models utilizing paired datasets for image denoising, which proves to be expensive and time-consuming. Current self-supervised unsupervised approaches typically rely blind-spot networks or sub-image pairs sampling, resulting in pixel information loss destruction of detailed structural information, thereby significantly constraining the efficacy such methods. In this paper, we introduce Prompt-SID, a prompt-learning-based single...
Many studies have concentrated on constructing supervised models utilizing paired datasets for image denoising, which proves to be expensive and time-consuming. Current self-supervised unsupervised approaches typically rely blind-spot networks or sub-image pairs sampling, resulting in pixel information loss destruction of detailed structural information, thereby significantly constraining the efficacy such methods. In this paper, we introduce Prompt-SID, a prompt-learning-based single...
Self-supervised video denoising aims to remove noise from videos without relying on ground truth data, leveraging the itself recover clean frames. Existing methods often rely simplistic feature stacking or apply optical flow thorough analysis. This results in suboptimal utilization of both inter-frame and intra-frame information, it also neglects potential alignment under self-supervised conditions, leading biased insufficient outcomes. To this end, we first explore practicality setting...
The extraction of auto-correlation in images has shown great potential deep learning networks, such as the self-attention mechanism channel domain and self-similarity spatial domain. However, realization above mechanisms mostly requires complicated module stacking a large number convolution calculations, which inevitably increases model complexity memory cost. Therefore, we propose pseudo 3D network (P3AN) to explore more efficient way capturing contextual information image de-noising. On...
For magnetic resonance (MR) images sharing visual characteristics, the internal structure repetitions of different scales are considerable image-specific priors. Following traditional algorithms, we try to combine external dataset-driven learning with self-similarity for MR image super-resolution (SR). We propose a pyramid orthogonal attention network (POAN) based on dual self-similarity. On one hand, by combining point-similarity and pyramid-similarity, sufficient spatial autocorrelation is...
The single-image super-resolution (SISR) network based on deep learning is dedicated to the mapping between low-resolution (LR) images and high-resolution (HR) images. optimal parameters of these networks often require extensive training large-scale external image databases. For medical magnetic resonance (MR) images, there a lack large data sets containing high-quality Some that perform well natural cannot be fully trained MR which limits (SR) performance. In traditional methods, non-local...
Abstract Fluorescence imaging with high signal-to-noise ratios has become the foundation of accurate visualization and analysis biological phenomena. However, inevitable photon shot noise poses a formidable challenge on sensitivity. In this paper, we provide spatial redundancy denoising transformer (SRDTrans) to remove from fluorescence images in self-supervised manner. First, sampling strategy based is proposed extract adjacent orthogonal training pairs, which eliminates dependence speed....
The feedback mechanism in the human visual system extracts high-level semantics from noisy scenes. It then guides low-level noise removal, which has not been fully explored image denoising networks based on deep learning. commonly used fully-supervised network optimizes parameters through paired training data. However, unpaired images without noise-free labels are ubiquitous real world. Therefore, we proposed a multi-scale selective (MSFN) with dual loss. We allow shallow layers to access...
How to properly model the inter-frame relation within video sequence is an important but unsolved challenge for restoration (VR). In this work, we propose unsupervised flow-aligned sequence-to-sequence (S2SVR) address problem. On one hand, model, which has proven capable of modeling in field natural language processing, explored first time VR. Optimized serialization shows potential capturing long-range dependencies among frames. other equip with optical flow estimator maximize its...
Self-supervised video denoising aims to remove noise from videos without relying on ground truth data, leveraging the itself recover clean frames. Existing methods often rely simplistic feature stacking or apply optical flow thorough analysis. This results in suboptimal utilization of both inter-frame and intra-frame information, it also neglects potential alignment under self-supervised conditions, leading biased insufficient outcomes. To this end, we first explore practicality setting...
Abstract A fundamental challenge in fluorescence microscopy is the inherent photon shot noise caused by inevitable stochasticity of detection. Noise increases measurement uncertainty, degrades image quality, and limits imaging resolution, speed, sensitivity. To achieve high-sensitivity beyond shot-noise limit, we provide DeepCAD-RT, a versatile self-supervised method for effective suppression time-lapse imaging. We made comprehensive optimizations to reduce its data dependency, processing...
During magnetic resonance imaging (MRI), the strong response to signal is usually displayed as structural edges and textures, which important for distinguishing different tissues lesions. In current superresolution (SR) methods with usage of deep learning, some low-level information tends gradually disappear network deepens, resulting in excessive smoothness high-frequency regions. This phenomenon particularly noticeable MRI poor brightness contrast small gray dynamic range. Although...