A5100692595- Advanced Neural Network Applications
- Domain Adaptation and Few-Shot Learning
- Advanced Image and Video Retrieval Techniques
- Multimodal Machine Learning Applications
- Robotics and Sensor-Based Localization
- Advanced Computational Techniques and Applications
- Human Pose and Action Recognition
- Brain Tumor Detection and Classification
- Machine Learning and ELM
- Ferroelectric and Negative Capacitance Devices
- COVID-19 diagnosis using AI
- Medical Image Segmentation Techniques
- Advanced Decision-Making Techniques
- Remote-Sensing Image Classification
- Advanced Image Processing Techniques
- Distributed Control Multi-Agent Systems
- Healthcare Operations and Scheduling Optimization
- Parallel Computing and Optimization Techniques
- Advanced Vision and Imaging
- Satellite Communication Systems
- Macrophage Migration Inhibitory Factor
- Hand Gesture Recognition Systems
- Time Series Analysis and Forecasting
- Advanced Data Storage Technologies
- Polynomial and algebraic computation
Xidian University
2021-2024
Guilin University of Aerospace Technology
2024
Hong Kong Science and Technology Parks Corporation
2024
Xihua University
2023
Institute of Microelectronics
2022
Chinese Academy of Sciences
2022
University of Chinese Academy of Sciences
2022
Peking University
2022
Beihang University
2011-2021
Beijing Institute of Technology
2011-2020
Modern machine learning datasets can have biases for certain representations that are leveraged by algorithms to achieve high performance without solve the underlying task. This problem is referred as “representation bias”. The question of how reduce representation a dataset investigated and new REPresentAtion bIas Removal (REPAIR) procedure proposed. formulates bias minimization an optimization problem, seeking weight distribution penalizes examples easy classifier built on given feature...
Optical and SAR image registration is important for multi-modal remote sensing information fusion. Recently, deep matching networks have shown better performances than traditional methods on matching. However, due to significant differences between optical images, the of existing learning still need be further improved. This paper proposes a self-distillation feature network (SDNet) registration, improving performance from structure optimization. Firstly, we explore impact different...
Unsupervised domain adaptation (UDA) carries out knowledge transfer from the labeled source to unlabeled target domain. Existing feature alignment methods in UDA semantic segmentation achieve this goal by aligning distribution between domains. However, these ignore domain-specific of In consequence, 1) correlation among pixels is not explored; and 2) classifier explicitly designed for distribution. To conquer obstacles, we propose a novel cluster framework, which mines when performing...
Deep convolutional neural networks attract increasing attention in image patch matching. However, most of them rely on a single similarity learning model, such as feature distance and the correlation concatenated features. Their performances will degenerate due to complex relation between matching patches caused by various imagery changes. To tackle this challenge, we propose multi-relation network (MRAN) for Specifically, fuse multiple relations (MR) matching, which can benefit from...
The success of deep learning heavily relies on large-scale data with comprehensive labels, which is more expensive and time-consuming to fetch in 3D compared 2D images or natural languages. This promotes the potential utilizing models pretrained than as teachers for cross-modal knowledge transferring. In this paper, we revisit masked modeling a unified fashion distillation, show that foundational Transformers languages can help self-supervised representation through training Autoencoders...
Deep convolutional networks are powerful for local feature learning and have shown advantages in image matching registration. However, the significant differences between cross-modal images increase challenge of The deep network should extract modality-invariant features to identify samples discriminative separate nonmatching samples. can invariant modality changes by multiple nonlinear mapping layers. it does not inevitably lose rich details affect discrimination features, degrading...
Recent state-of-the-art source-free domain adaptation (SFDA) methods have focused on learning meaningful cluster structures in the feature space, which succeeded adapting knowledge from source to unlabeled target without accessing private data. However, existing rely pseudo-labels generated by models that can be noisy due shift. In this paper, we study SFDA perspective of with label noise (LLN). Unlike conventional LLN scenario, prove follows a different distribution assumption. We also such...
Image registration and change detection are crucial for multitemporal remote sensing image analysis. The images should be registered before the information detection. Existing deep learning methods have shown significant advantages in tasks. They usually design two independent task-specific networks detection, respectively. These will learn from scratch rely on many labeled training datasets. This article finds that similar mechanisms, which focus extracting discriminative features. Inspired...
Artificial neural networks have acquired remarkable achievements in the field of artificial intelligence. However, it suffers from catastrophic forgetting when dealing with continual learning problems, i.e., loss previously learned knowledge upon new information. Although several algorithms been proposed, remains a challenge to implement these efficiently on conventional digital systems due physical separation between memory and processing units. Herein, software–hardware codesigned...
In this work, we tackle the challenging problem of denoising hand-object interactions (HOI). Given an erroneous interaction sequence, objective is to refine incorrect hand trajectory remove artifacts for a perceptually realistic sequence. This challenge involves intricate noise, including unnatural poses and relations, alongside necessity robust generalization new diverse noise patterns. We those challenges through novel approach, GeneOH Diffusion, incorporating two key designs: innovative...
Accurate segmentation of the hippocampal and its subfields from brain magnetic resonance imaging (MRI), which is a prerequisite for volume measurement, plays significant role in clinical diagnosis treatment many neurodegenerative diseases. It great significance precise hippocampus sub-regions.In this paper, we proposed approach based on support vector machine (SVM) combined 3D convolutional neural network (3D CNN) generative adversarial (GAN). In CNN-SVM model, representative features...
The contradiction that imaging system can acquire huge amount of image data while communication deliver only a very small part them has become bottleneck for the small-satellite based earth observation. In this paper, novel onboard selection strategy is designed to select most informative images are acquired by transmission. Specifically, worst reconstructed previously transmitted images, instead instantly image, selected since it possesses distinguishing information images. Experiment on...
Due to the different imaging mechanisms, there is a significant non-line difference between multi-modal images, which brings difficulties image registration. The traditional methods based on grayscale and handcraft features are difficult with obtain common source images. performances of deep local matching rely quality quantity detected keypoints, can be quite time-consuming register To achieve fast accurate registration, we propose global feature-based template method (GFTM) uses...
In order to improve the utilization ratio of surgery resources and develop a better scheduling method, we build MIP based model solve it effectively. Two numerical examples with real data are employed prove effectiveness model.
To find the best memory system for emerging workloads, traces are obtained during application's execution, then caches with different configurations simulated using these traces. Since program can be several gigabytes, simulation of cache performance is a time consuming process. Compute unified device architecture (CUDA) software development platform which enables programmers to accelerate general-purpose applications on graphics processing unit (GPU). This paper presents real multi-core...
This article focuses on end-to-end image matching through joint key-point detection and descriptor extraction. To find repeatable high discrimination key points, we improve the deep network from perspectives of structure optimization. First, propose a concurrent multiscale detector (CS-det) network, which consists several parallel convolutional networks to extract features multilevel discriminative information for detection. Moreover, introduce an attention module fuse response maps various...
Continual Learning Artificial neural networks suffer from catastrophic forgetting when meeting sequential tasks. In article number 2200026, Zhongrui Wang, Dashan Shang, and co-workers propose a metaplasticity-inspired mixed-precision continual learning model to address this issue. By deploying it on neuromorphic prototype system with the in-memory computing paradigm, outperforming energy efficiency high accuracy are demonstrated, paving promising way for autonomous edge systems.