Unsupervised source-free domain adaptation approaches focus on transferring knowledge from a labeled source domain to an unlabeled target domain, where access to source data is unavailable during the target adaptation process to further enhance data privacy protection. Existing methods generate pseudo labels to target samples based on pre-trained source model to iteratively optimize target model. However, inaccurate pseudo labels can negatively impact prediction results in target domain. Therefore, we introduce a novel framework with diffusion model-guided feature transformation to effectively train the target model. Instead of generating synthetic samples, our method leverages the diffusion model to refine target features, transforming them to better resemble the source domain feature space. To further enhance alignment, we introduce archetype feature representing each source class and design a contrastive learning scheme to promote compact intra-class and separated inter-class representations. Extensive experiments on some benchmark datasets validate the effectiveness and superiority of the proposed method.
Accurate swallowing assessment via surface electromyography (sEMG) is hindered by limited channel interaction modeling, uniform temporal processing of event-driven signals, and scarce labeled clinical data. We propose SwallowMamba, a hybrid framework that integrates multi-scale convolutional feature extraction with a multi-channel cooperative scanning state space model (MCCS-SSM). MCCS-SSM assigns channel-specific state space parameters and employs a time-varying dynamic cooperation matrix via scaled dot-product attention to model inter-muscle coordination at each time step. An event-aware adaptive step-size (EAAS) mechanism modulates temporal resolution by integrating short-time energy, spectral flux, and instantaneous amplitude variation, allocating finer discretization to swallowing events and coarser steps to rest. To address labeled data scarcity, a hierarchical VICReg (H-VICReg) self-supervised pretraining strategy leverages the large-scale NinaPro hand sEMG dataset for cross-anatomical transfer with multi-layer regularization. Evaluated on a five-class swallowing sEMG dataset (11 subjects, including 3 laryngeal cancer patients) using leave-one-subject-out cross-validation, SwallowMamba achieves 93.5
Time series forecasting (TSF) is critical in various real-world applications, and has been extensively studied in the research community. Recent state-of-the-art (SOTA) methods for this task largely rely on fine-tuning or prompt engineering within the context of large language models (LLMs). However, fine-tuning can compromise the original language capabilities of the models, thereby affecting their generalization. Meanwhile, incorporating prompts significantly increases computational overhead. To address these challenges, we propose Frozen-Time, a simple, lightweight, and efficient approach leveraging LLMs for effective TSF across various domains. Technically, Frozen-Time first divides the input time series into non-overlapping patches with uniform temporal resolution. This naturally facilitates the capture of local semantics and the modeling of internal correlations within the time series. After the divided patches pass through the adaptive input embedding linear layer, we propose a frozen module as the forecaster, which consists of a single frozen transformer layer from the LLM. Finally, the forecasting results are generated through a lightweight output projection layer. This adaptive design avoids the overlapping time series misinterpretation, significantly reduces computational costs, and maintains high performance. Extensive experiments on popular public datasets demonstrate the robustness, computational efficiency, and superior generalization capabilities of Frozen-Time, consistently outperforming SOTA methods in long-term, short-term, few-shot, and zero-shot forecasting tasks. In addition, our method significantly reduces parameters and computational requirements, making it feasible for deployment on less advanced GPUs.
Pixel-level infrared and visible image fusion (pixel-level IVIF) integrates multimodal imaging features to generate enhanced images that simultaneously preserve thermal radiation saliency and fine texture details. This technique has shown great potential in various applications such as night vision enhancement, remote sensing, and autonomous driving. In recent years, a growing number of deep learning-based pixel-level IVIF methods have been proposed. This paper aims to help researchers in the field by providing a comprehensive review of the latest advances in pixel-level IVIF based on deep learning. We review over 120 representative papers, covering the background, taxonomies, development characteristics, benchmark datasets, and performance comparisons of the reviewed methods. In addition, we discuss the current challenges and potential future directions in pixel-level IVIF research. This review serves as an up-to-date reference for researchers interested in pixel-level IVIF and those looking to apply fusion techniques to broader application scenarios.
Person search presents unique challenges by requiring simultaneous pedestrian detection and identification within unified frameworks. While anchor-free models have emerged to address computational inefficiencies from dense object anchors, their simplified architectures often suffer from object conflicts and region misalignment issues. To resolve these limitations, we introduce CAPS, a novel contextually aligned end-to-end anchor-free person search framework that explores more contextual cues in scene feature for identification. Our architecture preserves spatial-contextual information through dedicated identity branches that explicitly align features along channel and spatial dimensions. We further introduce region-aligned knowledge distillation to ensure spatial robustness, transferring the teacher model’s discriminative capabilities via multi-level adapters. This strategy decouples intensive training supervision from efficient deployment inference, achieving an optimal balance between performance and speed. Extensive experiments demonstrate state-of-the-art performance of CAPS on CUHK-SYSU and PRW benchmarks and validate our approach’s simultaneous optimization of accuracy, computational efficiency, and architectural simplicity.
