夏靖远

This paper introduces C²SSM, a novel visual state space model for ultra-high-definition (UHD) image restoration. Unlike existing methods that scan every pixel, C²SSM adopts a cluster-centric paradigm: it first aggregates pixels into semantic clusters, then performs global modeling only on the cluster centroids, and finally diffuses the context back to all pixels. This significantly reduces computational cost while preserving high-quality restoration. Experiments on five UHD tasks show state-of-the-art performance with much lower complexity.

WST-DRFSL is a semi-supervised few-shot target recognition framework that employs a dynamic refinement strategy. This strategy alternately refines the classifier and the entire model, accelerating convergence while achieving high recognition accuracy across extensive datasets.

HiFormer is a dual-branch Transformer architecture designed for Ultra-High-Definition (UHD) image restoration. It combines a high-resolution branch that preserves fine details using directionally-sensitive large-kernel convolutions with a low-resolution branch that models global context via self-attention. This collaboration effectively compensates for high-frequency losses caused by downsampling and attention mechanisms, enabling efficient and high-fidelity restoration on consumer-grade GPUs.

PATK reformulates time series tokenization as a physics-aware semantic segmentation process over dual-domain signal distributions, leveraging a hidden Markov modeling mechanism to dynamically adapt token boundaries and achieve robust, interpretable representation learning across diverse temporal domains.

LAOF is a general unsupervised radar imaging framework. We have verified the SOTA performance of LAOF in sparse aperture ISAR imaging and auto-focusing problems under extremely low sparsity rates, extremely low signal-to-noise ratios, and in real-world data.

LASQ reformulates low-light image enhancement as a statistical sampling process over hierarchical luminance distributions, leveraging a diffusion-based forward process to autonomously model luminance transitions and achieve unsupervised, generalizable light restoration across diverse illumination conditions.

SAKE performs unsupervised HSI super-resolution via adaptive blur kernel estimation, requiring no extra data and demonstrating superior generalization across scenarios.

SAIG introduces a semantic-aware generation framework using segmentation-based refinement to produce high-fidelity ISAR images from optical inputs.

AORC learns attitude-aware latent representations and transforms them into high-fidelity ISAR images from optical inputs through a Brownian-Bridge-based diffusion process.

STC leverages SIFT features and decoupled learning to tackle class imbalance in SAR target recognition, achieving strong generalization and SOTA performance.

The proposed DKP is a plug-and-play kernel estimation tool, which can be combined with the off-the-shelf image restoration model, e.g., DIP and Diffusion model, to realize unsupervised blind image super-resolution.

DDSR is an unsupervised blind SISR method, which can handle different degradations, such as partial occlusion, noise, and low-light conditions.

This paper proposes a plug-and-play blind image super-resolution method based on the Meta-learning and Markov Chain Monte Carlo, which contributes to preventing bad local optimal solutions from the optimization perspective.

The proposed MLDP aims to realize the image translation from optical domain to ISAR domain. To generate realistic ISAR images, MLDP learns the scattering distribution features and the classification identifying features in physical and task perspectives. Meanwhile, MLDP applies the meta-learning strategy to improve its generalization ability to generate ISAR images under limited training samples.

This paper proposes a meta-learning-based alternating minimization (MLAM) method for nonconvex problems of multiple variables, which aims to minimize a part of the global losses over iterations instead of carrying minimization on each subproblem, and it tends to learn an adaptive strategy to replace the handcrafted counterpart resulting in advance on superior performance.

An improved Localized incomplete multiple kernel k-means (LI-MKKM), called LI-MKKM with matrix-induced regularization (LI-MKKM-MR), is proposed by incorporating a matrix-induced regularization term to handle the correlation among base kernels. We theoretically show that the local kernel alignment is a special case of its global one with normalizing each base kernel matrices.

This letter proposes a learning aided gradient descent (LAGD) algorithm to solve the weighted sum rate (WSR) maximization problem for multiple-input single-output (MISO) beamforming. The proposed LAGD algorithm directly optimizes the transmit precoder through implicit gradient descent based iterations, at each of which the optimization strategy is determined by a neural network, and thus, is dynamic and adaptive.