Point-supervised temporal action localization (P-TAL) aims to localize actions in untrimmed videos with only one action-level timestamp annotation. Owing to label sparsity in P-TAL, most existing methods locate action by boundary-free considering classification, which results in incomplete action localization. In this paper, we propose a novel point-to-boundary hierarchical learning framework that generates dense and reliable pseudo-action proposals to provide complete action annotation containing boundaries for a multi-granularity localization model with single-stage inference. Concretely, we first achieve action proposal boundary enhancement which takes the annotated points as corresponding action centers to search for the boundary-accurate action proposals by fitting a Gaussian prior according to Class Activation Sequences. Furthermore, to more noise-resiliently learn boundary information of actions from the obtained proposals, we introduce a hierarchical boundary regression algorithm and a multi-level attention loss function. Extensive experimental results on three challenging benchmarks demonstrate the state-of-the-art performance and the generalization of the proposed framework. Notably, our method even shows faster single-stage inference with better performance compared to the existing methods.

Emotion recognition using electroencephalography (EEG) has emerged as a key noninvasive physiological modality for human-computer interaction, intelligent healthcare, and mental-health monitoring. However, existing approaches often treat temporal and spatial features of EEG signals in isolation and lack bidirectional, deep interactive fusion mechanisms. Furthermore, multiscale analysis of temporal and spatial information remains limited in scope. To overcome these challenges, we introduce a Dual-Branch Multiscale Bi-Mamba framework (DBM-BiMamba). DBM-BiMamba is a unified, multiscale, bidirectional, spatiotemporal fusion framework. It comprises three components that are optimized together: a Multiscale Temporal Feature Learning (MTFL) branch, a Hierarchical Spatial Feature Learning (HSFL) branch and a Bidirectional Fusion Mamba (Bi-Mamba) module. The MTFL branch uses parallel depthwise separable convolutions, followed by a transformer encoder, to capture local micro-dynamics and long-range temporal dependencies. The HSFL branch performs hierarchical graph convolutions with node-wise attention to emphasize critical inter-electrode relationships and produce discriminative topological embeddings. The Bi-Mamba module then applies forward and backward state-space modeling to the temporal and spatial embeddings, fusing them at the sequence level and enabling efficient bidirectional spatiotemporal interactions rather than simply concatenating the features. Extensive ten-fold subject-independent cross-validation in the DEAP, SEED, and SEED-IV datasets demonstrates stateof-the-art accuracy of 95.55% for arousal and 95.02% for valence in DEAP, 94.77% in SEED, and 89.68% in SEED-IV. Ablation studies confirm the complementary gains of each component.

Temporal action segmentation (TAS) aims to classify the action category of each frame in untrimmed videos, with RGB videos or skeleton sequences as common inputs. Most existing methods follow a two-stage pipeline of feature extraction and temporal modeling. However, we observe two key limitations in their spatio-temporal modeling: (i) Existing temporal modeling modules conduct frame-level and action-level interactions at a single fixed temporal resolution, which over-smooths temporal features and blurs action boundaries; (ii) Skeleton-based methods generally adopt temporal modeling modules originally designed for RGB video data, causing a misalignment between extracted features and temporal modeling modules. To address these issues, we propose a novel Interaction-based framework for Action segmentation (InterAct). Firstly, we introduce multi-scale frame-action interaction (MFAI) to facilitate frame-action interactions across varying temporal scales. This enhances the model’s ability to capture complex temporal dynamics, producing more expressive temporal representations and alleviating the over-smoothing issue. Meanwhile, recognizing the complementary nature of different spatial modalities, we further propose decoupled spatial modality interaction (DSMI). It decouples the modeling of spatial modalities and applies a deep fusion strategy to interactively integrate multi-scale spatial features. This results in more discriminative representations that better support temporal modeling. Extensive experiments on seven large-scale benchmarks show that InterAct significantly outperforms state-of-the-art methods on both RGB-based and skeleton-based TAS. Further evaluations on multi-person interactions and simulated occlusions demonstrate improved robustness under imperfect skeleton observations. The code is available at https://github.com/gdxxu/InterAct-main.

Timestamp-supervised temporal action segmentation (TS-TAS) has attracted extensive attention due to its low annotation cost, which only requires a single timestamp label per action segment. However, the sparsity of supervision causes significant performance bottlenecks, especially in ambiguous action boundaries and visually similar categories. Existing methods focus on video modality alone and ignore the rich semantics in text annotations, which makes it difficult to overcome the performance bottleneck caused by sparse supervision. To address these challenges, this paper proposes an adaptive routing network based on multimodal fusion for TS-TAS. The model aligns visual and text features semantically in a unified embedding space, dynamically enhances discriminative features via an adaptive routing module, and generates pseudo-text features to solve the text absence problem during inference. Extensive experiments on GTEA, 50Salads, and Breakfast datasets show that our method outperforms state-of-the-art TS-TAS approaches and achieves competitive performance with fully supervised methods at a much lower annotation cost. Qualitative and quantitative results validate the effectiveness of the proposed modules in improving segmentation accuracy, boundary localization, and the discrimination of similar action categories.

Since underwater videos involve scarce paired data and mismatched physical priors, underwater visual enhancement faces significant challenges in hybrid static-dynamic scenes. Existing methods typically design single-frame enhancement paradigms to address cross-frame optical variations, resulting in temporally inconsistent video and redundant computations. Although spatial-temporal feature fusion is a common inter-frame correlation strategy to mitigate these issues, mainstream fusion strategies struggle to handle cross-frame variations caused by non-uniform channel attenuation and regional degradation. Therefore, we propose a physics-semantic collaborative dual-branch framework. This framework employs image-branch priors as unified enhancement prompts, guiding the video branch to construct more robust semantic understanding and physical estimation across multiple frames. Specifically, we incorporate a selective channel fusion mechanism into Mamba to efficiently aggregate complementary spatial-temporal cues. Moreover, we devise a novel physical parameter estimation paradigm for adaptively interpreting dynamic degradation, thereby guiding color correction and degradation removal in underwater video. Extensive experiments on six datasets demonstrate that the proposed method achieves superior enhancement performance compared to the state-of-the-art methods.

To address the issues of poor handling of ground object change boundary details in complex scenes and insufficient synergy caused by the decoupling of semantic and change subtasks in existing semantic change detection methods for remote sensing images, this paper proposes an improved algorithm based on contrastive learning. A Cross-Layer Attention Fusion FPN Decoder is designed. Through a multi-stage feature fusion strategy, it integrates low-level spatial details and high-level contextual information to generate more discriminative feature representations. Additionally, a Pixel-Level Change Modeling with Classification Features method is proposed, which shifts change modeling from the decision layer to the feature fusion stage. A dynamic change-aware loss is introduced to establish a collaborative training mechanism between the semantic segmentation and change detection tasks, enhancing the model’s sensitivity to change regions. In contrastive learning, a boundary hard sample mining strategy is proposed to prioritize the optimization of category boundary pixels, strengthening the model’s ability to recognize complex boundaries. The method is validated on two public datasets. Experimental results show that, compared with existing methods, the proposed method achieves significant improvements on multiple evaluation metrics for the tasks of change detection and semantic segmentation.

Overcoming visual degradation in challenging imaging scenarios is essential for accurate scene understanding. Although deep learning methods have integrated various perceptual capabilities and achieved remarkable progress, their high computational cost limits practical deployment under resource-constrained conditions. Moreover, when confronted with diverse degradation types, existing methods often fail to effectively model the inconsistent attenuation across color channels and spatial regions. To tackle these challenges, we propose DWMamba, a degradation-aware and weight-efficient Mamba network for image quality enhancement. Specifically, DWMamba introduces an Adaptive State Space Module (ASSM) that employs a dual-stream channel monitoring mechanism and a soft fusion strategy to capture global dependencies. With linear computational complexity, ASSM strengthens the models ability to address non-uniform degradations. In addition, by leveraging explicit edge priors and region partitioning as guidance, we design a Structure-guided Residual Fusion (SGRF) module to selectively fuse shallow and deep features, thereby restoring degraded details and enhancing low-light textures. Extensive experiments demonstrate that the proposed network delivers superior qualitative and quantitative performance, with strong generalization to diverse extreme lighting conditions. The code is available at https://github.com/WindySprint/DWMamba.

Accurate image registration is essential for various remote sensing applications, particularly in multi-temporal image analysis. This paper introduces DVF-NET, a novel deep learning-based framework for dual-temporal remote sensing image registration. DVF-NET integrates two displacement vector fields to address nonlinear distortions caused by significant variations between images, enabling more precise image alignment. A key innovation of this method is the incorporation of a Structural Attention Module (SAT), which enhances the model’s ability to focus on structural features, improving the feature extraction process. Additionally, we propose a novel loss function design that combines multiple similarity metrics, ensuring more comprehensive supervision during training. Experimental results on various remote sensing datasets indicate that the proposed DVF-NET outperforms the existing methods in both accuracy and robustness, particularly when handling images with substantial geometric distortions such as tilted buildings. The results validate the effectiveness of our approach and highlight its potential for various remote sensing tasks, including change detection, land cover classification, and environmental monitoring. DVF-NET provides a promising direction for the advancement of remote sensing image registration techniques, offering both high precision and robustness in complex real-world scenarios.

Mainstream spectral reconstruction methods typically meticulously design complex and computationally intensive architectures in convolutional neural networks (CNNs) or Transformers to model the mapping from RGB to hyperspectral image (HSI). However, the bottleneck in achieving accurate spectral reconstruction may not lie in model complexity. Direct end-to-end learning on limited training samples struggles to encapsulate discriminative and generalizable feature representations, leading to overfitting and consequently suboptimal reconstruction fidelity. To address these challenges, we propose a new Masked Autoencoder-based Knowledge Transfer network for Spectral Reconstruction from RGB images (MAE-KTSR). MAE-KTSR decouples the feature representation process into a two-stage paradigm, facilitating a holistic comprehension of diverse objects and scenes, thereby enhancing the generalizability of spectral reconstruction. In the first stage, we introduce Spatial-Spectral Masked Autoencoders (S-2-MAE) to extract discriminative spectral features through masked modeling under constrained spectral conditions. S-2-MAE reconstructs spectral images from partially masked inputs, learning a generalizable feature representation that provides useful prior knowledge for RGB-to-HSI reconstruction. In the second stage, a lightweight convolutional reconstruction network is deployed to further extract and aggregate local spectral-spatial features. Specifically, an Inter-Stage Feature Fusion module (ISFF) is introduced to effectively exploit the global MAE-based spectral priors learned in the first stage. Experimental results on three spectral reconstruction benchmarks (NTIRE2020-Clean, CAVE, and Harvard) and one real-world hyperspecral dataset (Pavia University) demonstrate the effectiveness of MAE-KTSR. Additionally, MAE-KTSR is experimentally validated to facilitate downstream real-world applications, such as HSI classification.

Multi-modal Emotion Recognition (MER) has demonstrated competitive performance in affective computing, owing to synthesizing information from diverse modalities. However, many existing approaches still face unresolved challenges, such as: (i) how to learn compact yet representative features from multi-modal data simultaneously, and (ii) how to address differences among subjects and enhance the generalization of the emotion recognition model, given the diverse nature of individual biological signals. To this end, we propose a Dynamic Interactive Network with Self-Distillation (DISD-Net) for cross-subject MER. The DISD-Net incorporates a dynamin interactive module to capture the intra- and inter-modal interactions from multi-modal data. Additionally, to enhance compactness in modal representations, we leverage the soft labels generated by the DISD-Net model as supplemental training guidance. This involves incorporating self-distillation, aiming to transfer the knowledge that the DISD-Net model contains hard and soft labels to each modality. Finally, domain adaptation (DA) is seamlessly integrated into the dynamic interactive and self-distillation components, forming a unified framework to extract subject-invariant multi-modal emotional features. Experimental results indicate that the proposed model achieves a mean accuracy of 75.00% with a standard deviation of 7.68% for the DEAP dataset and a mean accuracy of 65.65% with a standard deviation of 5.08% for the SEED-IV dataset.

Water quality prediction and monitoring are crucial for environmental protection. This study proposes an improved long short-term memory neural network model for complex time-series water quality data. The model optimizes traditional long short-term memory structures to address the fluidity of water quality data. Additionally, empirical mode decomposition is introduced to capture water pollution characteristics and identify trends and fluctuations. Simulation results indicate that the optimal configuration included a sliding window size of 4 and 20 hidden layer nodes, converging after 22 training iterations with a loss value of approximately 0.027. The improved model achieved a 31

As an economically feasible approach for hyperspectral image (HSI) super-resolution, fusing HSI with multispectral image (MSI) utilizes the complementary nature of cross-modality information. Given the common presence of repetitive textures and structures, non-local self-similarity are widely used in natural image super-resolution as an efficient prior information. However, self-similarity in HSI has not been fully explored. Existing methods primarily focus on spatial dimension self-similarity, ignoring the role of cross-band self-similarity and spectral-spatio correlations in HSI, resulting in sub-optimal solutions. In this work, we heuristically propose a new framework (S3-Net) to learn the self-similarity prior across spectral-spatio dimensions comprehensively for reconstructing high-resolution HSI. Specifically, the proposed S3-Net is built upon multiple designed Tribranch Self-Similarity Fusion (TSSF) blocks. Each block features three parallel branches which are responsible for capturing the complex interactions between the spatial and spectral dimensions of the input tensor, and utilizing permutation operations to facilitate this multi-dimensional analysis. We further design and insert a Non-Local Self-Similarity Attention (NLSSA) module for each branch of the TSSF block which efficiently aggregate globally relevant correlation features with minimal computational overhead. Extensive experiments on four HSI datasets, including Houston, Pavia Center, Urban, and Xiongan, demonstrated the superior performance and effectiveness of S3-Net. Code is available at https://github.com/wxy11-27/S3-Net.

Mainstream approaches to spectral reconstruction primarily focus on Convolution- and Transformer-based architectures. However, CNN methods fall short in handling long-range dependencies, whereas Transformers are constrained by computational efficiency limitations. Therefore, constructing a efficient spectral reconstruction network while ensuring the quality of reconstructed hyperspectral images (HSIs) has become a major challenge. Recent breakthroughs in the state-space model (e.g., Mamba) have attracted significant attention from natural language processing to vision tasks due to its near-linear computational efficiency and superior performance, prompting our investigation into its potential for spectral reconstruction problems. To this end, we introduce the Gradient-integrated Mamba for Spectral Reconstruction from RGB Images, dubbed GMSR-Net. GMSR-Net is a lightweight model characterized by a global receptive field and linear computational complexity. Its core comprises multiple stacked Gradient Mamba (GM) blocks, each featuring a tri-branch structure. Building upon the efficient global feature representation from the Mamba, we further innovatively propose spatial gradient attention and spectral gradient attention to guide the reconstruction of spatial and spectral cues. GMSR-Net demonstrates a significant accuracy-efficiency trade-off, achieving state-of-the-art performance while markedly reducing the number of parameters and computational burdens. Compared to existing approaches, GMSR-Net slashes parameters and FLOPs by substantial margins of 8 times and 20 times, respectively. Code is available at https://github.com/wxy11-27/GMSR.

Temporal distribution shifts occur not only in low-dimensional time-series data but also in high-dimensional data like videos. This phenomenon leads to significant performance degeneration in video understanding methods such as streaming temporal action segmentation. To address this issue, we propose a flexible streaming temporal action segmentation model with diffusion models (FSTAS-DM). By utilizing streaming video clips with varying feature distributions as control conditions, our model can adapt to the shifts and inconsistency of the distribution between the training and testing domains. Additionally, we have introduced a multistage conditional control training strategy (MSCC), which enhances the temporal generalization ability of the model. Our method demonstrates commendable performance on datasets like GTEA, 50Salads, and Breakfast.

Recently, deep learning-based remote sensing image super-resolution (RSISR) techniques have achieved significant progress, but challenges remain in preserving critical edge details essential for high-quality image reconstruction. These details are crucial for tasks such as object recognition, change detection, and accurate analysis in remote sensing imagery. Furthermore, existing RSISR methods typically require substantial computational resources, making them unsuitable for resource-constrained edge devices. To address these challenges, we propose a novel edge-guided super-resolution network (EGSRN). The network employs an edge extraction module (Edge Net) to explicitly extract edge information from low-resolution (LR) images, combined with multilayer feature extraction modules (FEMs) and an edge information fusion (EIF) mechanism to progressively integrate edge and image features. This design enables precise recovery of edge details, significantly enhancing the overall visual quality of the reconstructed images. Edge-aware processing enhances visual fidelity while also improving the accuracy of downstream tasks, such as classification, object detection, and change analysis. Furthermore, the network incorporates lightweight designs such as depthwise separable convolutions and channel shuffling to effectively reduce computational demands. Comprehensive experiments were conducted on two remote sensing datasets, and the model's parameter count and floating-point operations (FLOPs) were evaluated. Results demonstrate that the proposed method achieves an excellent balance between performance and model complexity, delivering superior super-resolution reconstruction quality while maintaining low computational costs, making it well-suited for resource-limited real-world applications.

Recently, deep learning-based remote sensing image super-resolution (RSISR) has achieved remarkable progress. However, conventional super-resolution (SR) methods usually assume a fixed and known degradation process (e.g., bicubic downsampling), which often leads to significant performance degradation when applied to real-world data with diverse and unknown degradations. To overcome this limitation, we propose DCR-SRNet, a novel degradation-contrastive and wavelet-guided network for blind RSISR. The proposed network incorporates three key innovations: First, we design a contrastive degradation representation learning strategy that disentangles degradation priors from scene semantics by pulling together representations of identical degradations across different scenes while pushing apart those of different degradations within the same scene. Second, we introduce a wavelet-guided patch-wise weighted loss module, which employs wavelet decomposition and patch-level discrimination scores to adaptively reweight the pixel-wise loss, thereby enhancing the recovery of edge and texture details. Third, we design an adaptive modulation block (AMB) that injects degradation priors into the reconstruction process through feature- and channel-wise modulation, enabling robust adaptation to diverse degradations. Extensive experiments on three benchmark remote sensing datasets demonstrate that DCR-SRNet significantly outperforms state-of-the-art methods, particularly in preserving structural and textural details.

The underwater images often suffer from color deviations and blurred details. To address these issues, many methods employ networks with an encoder/decoder structure to enhance the images. However, the direct skip connection overlooks the differences between pre- and post-features, and deep network learning introduces information loss. This paper presents an underwater image enhancement network that focuses on pre-post differences. The network utilizes a multi-scale input and output framework to facilitate the underwater image enhancement process. A novel cross-wise transformer module (CTM) is introduced to guide the interactive learning of features from different periods, thereby enhancing the emphasis on detail-degraded regions. To compensate for the information loss within the deep network, a feature supplement module (FSM) is devised for each learning stage. FSM merges the multi-scale input features, effectively enhancing the visibility of underwater images. Experimental results across several datasets demonstrate that the integrated modules yield significant enhancements in network performance. The proposed network exhibits outstanding performance in both visual comparisons and quantitative metrics. Furthermore, the network also exhibits good adaptability in additional visual tasks without the need for parameter tuning. The code and results are released in https://github.com/WindySprint/UVZ.

Action segmentation in untrimmed videos is essential for comprehensive video understanding. Despite significant progress in unsupervised methods, capturing both long-range dependencies and short-duration actions simultaneously remains a challenging task. To address this challenge, this paper introduces the Cluster-Refined Optimal Transport (CROT) method, combining hierarchical clustering and optimal transport for unsupervised action segmentation. We first hierarchically cluster video frame representations to capture long-range dependencies and generate pseudo-boundaries. Initial pseudo-labels are then obtained via optimal transport, ensuring short-duration actions are recognized. Finally, these pseudo-labels are refined using the pseudo-boundaries, resulting in the final segmentation output. Extensive experiments on three public datasets, i.e., YouTube Instructions, Breakfast, and 50Salads, demonstrate that our method performs on par with or better than previous approaches.

The streaming temporal action segmentation (STAS) task, a supplementary task of temporal action segmentation (TAS), has not received adequate attention in the field of video understanding. Existing TAS methods are constrained to offline scenarios due to their heavy reliance on multimodal features and complete contextual information. The STAS task requires the model to classify each frame of the entire untrimmed video sequence clip by clip in time, thereby extending the applicability of TAS methods to online scenarios. However, directly applying existing TAS methods to SATS tasks results in significantly poor segmentation outcomes. In this article, we thoroughly analyze the fundamental differences between STAS tasks and TAS tasks, attributing the severe performance degradation when transferring models to model bias and optimization dilemmas. We introduce an end-to-end streaming video TAS model with reinforcement learning (SVTAS-RL). The end-to-end modeling method mitigates the modeling bias introduced by the change in task nature and enhances the feasibility of online solutions. Reinforcement learning (RL) is utilized to alleviate the optimization dilemma. Through extensive experiments, the SVTAS-RL model significantly outperforms existing STAS models and achieves competitive performance to the state-of-the-art (SOTA) TAS model on multiple datasets under the same evaluation criteria, demonstrating notable advantages on the ultralong video dataset EGTEA. Our code is publicly available at https://github.com/Thinksky5124/SVTAS.

Recently, deep learning-based remote sensing image super-resolution (RSISR) methods have achieved remarkable progress. However, effectively preserving high-frequency details remains a significant challenge, as these features are critical for downstream tasks such as object detection, change analysis, and scene classification. Moreover, relying solely on the information contained in low-resolution (LR) images often results in the loss of structural details, thereby degrading reconstruction quality. To address these issues, we propose a novel wavelet-guided and feature-aware super-resolution network (WFA-SRNet). The proposed network adopts a dual-branch architecture, consisting of a feature extraction block (FEB) and a high-frequency extraction (HFE), to collaboratively model semantic structures and fine-grained textures. Specifically, FEB integrates a shift-window cross attention (SWCA) mechanism and a dictionary-based similarity matching strategy to capture nonlocal self-similarities, while the HFE branch incorporates a wavelet-domain high-frequency enhancement module (WD-HFE), which explicitly decomposes and reconstructs frequency components via discrete wavelet transform (DWT) and inverse DWT (IDWT) to enhance edge and texture recovery. Furthermore, a fusion attention (FA) module is designed to guide the integration of multisource features from both semantic and high-frequency pathways. Extensive experiments on multiple benchmark remote sensing datasets demonstrate that WFA-SRNet achieves superior reconstruction performance, particularly in restoring structural and textural details. Additionally, the proposed method significantly improves the accuracy of downstream classification tasks, showing strong potential for practical RSISR applications.