Automatic pulmonary lesion segmentation remains a formidable challenge due to the extreme morphological heterogeneity of pathological regions and the confounding influence of complex background topologies. To transcend these limitations, frequency-domain representations are introduced to exploit multi-spectral cues that are often inaccessible to purely spatial visual learning. However, existing frequency-based paradigms remain constrained within the unimodal visual domain, lacking the integration of high-level diagnostic semantics. Furthermore, the incorporation of diagnostic language introduces significant alignment difficulties: a substantial semantic-spatial gap exists between relation-based textual descriptors and grid-oriented visual features, and an inherent inconsistency persists between static textual guidance and hierarchical visual reconstructions. To address these critical challenges, a topology-aware vision-language framework, termed DFT-Seg, is proposed to explicitly bridge the multi-modal representation gap. Specifically, a prior-guided spatial grounding mechanism is established to resolve coordinate ambiguities by leveraging structural invariants. Moreover, to overcome unimodal constraints, a frequency modulation scheme is formulated under explicit textual guidance, utilizing semantic constraints to selectively amplify pathological edge responses while suppressing domain-specific noise. Finally, a progressive semantic refinement strategy is designed within the decoding phase to synchronize linguistic priors with hierarchical visual reconstructions. Extensive experiments on chest radiography and computed tomography datasets demonstrate that DFT-Seg achieves 91.41% and 84.17% in Dice coefficient and Intersection over Union on QaTa-COV19, along with 80.42% and 67.26% on MosMedData+, respectively, establishing consistent superiority over existing competitive baselines. The code is publicly available at https://github.com/NuistSMS/DFT_Seg-main.
Portrait style transfer has been widely applied in film, animation, and so on. However, current portrait style transfer models often suffer from the loss of facial structure during input image reconstruction and the generation of artifacts in the resulting stylized images. To address this, we propose an exemplar-based unpaired portrait style transfer with a dual-branch GAN framework. Firstly, we design two structurally distinct feature encoders to enhance feature extraction capabilities for both content and style images. In addition, to reduce structure loss during content image reconstruction, a hybrid latent space is introduced. Meanwhile, we introduce a portrait reconstruction loss to preserve overall contours. Finally, we construct a feature cross-fusion module (FCFM) to enhance the fusion of the features. This module not only adjusts the style of the fused features but also reduces potential artifacts from the fusion process. Qualitative and quantitative analyses demonstrate that our method outperforms baselines in generating stylized images across several styles.
The Visual Geometry Grounded Transformer (VGGT) demonstrated that an overcomplete prediction strategy, in which a single model infers multiple 3D attributes, significantly enhances performance. However, its reliance on a large-scale, private dataset has created a research bottleneck, hindering further development and fair comparison. This paper addresses this challenge with two key contributions. First, we propose a practical and reproducible pipeline to generate high-quality pseudo-ground truth data from public resources, enabling VGGT-based research in compute-restricted settings. Second, we enhance VGGT’s over-complete strategy by integrating a 3D Gaussian Splatting (3DGS) head. This introduces a complementary, region-level supervisory signal (image reconstruction loss) that synergizes VGGT’s existing geometric losses. Our experiments, conducted under identical annotation-sparse environments, show that our proposed model consistently outperforms the baseline VGGT in camera pose estimation and reconstruction quality. These results validate our hypothesis that this synergistic learning approach is highly effective, especially in low-resource settings, and provides a valuable pathway for developing more accessible and efficient 3D vision systems.
Multispectral object detection, fusing visible (RGB) and infrared (IR) modalities, is pivotal for robust all-weather perception. However, a critical phenomenon—Modality Failure—plagues existing methods, where naive fusion causes multimodal models to underperform unimodal baselines. We identify this bottleneck as stemming from two limitations: inadequate intra-modality representation and insufficient cross-modal interaction, which jointly impede the exploitation of complementarity. To mitigate this issue, we propose a Multiscale intra-Modal Feature Refinement and Local Deformable convolution Fusion (M2FRe-LDF) to reinforce feature representations while facilitating inter-modal information interaction to excavate more comprehensive complementary information. The framework comprises two synergistically collaborative components: the M2FRe module and the LDF module. Specifically, the M2FRe module dynamically assesses the significance of input features along the channel dimension to enhance the representation of critical information. Building upon these refined features, the LDF module then captures inter-modality correlations by leveraging deformable convolution to extract complementary information from a global perspective, thereby enabling a comprehensive and robust information interaction. Extensive experiments on three public datasets demonstrate the effectiveness of our method, achieving 65.1%, 39.7% and 55.9% mAP in LLVIP, FLIR and M3FD, effectively alleviating the phenomenon of Modality Failure.Our code is available at:https://github.com/yuyn133/MMFRe-LDF.
Artificial intelligence in low-sample-size scenarios offers new opportunities for preserving traditional martial arts. This study introduces a few-shot action recognition method using a graph convolutional neural network (GCN) to enhance martial arts accessibility. The approach extracts skeletal data from videos, via monocular 3D pose estimation, encodes skeletal sequences into a metric space using GCN, and employs metric learning for classification. Inspired by traditional martial arts principles, we integrate an interpretable attention module to highlight critical joint interactions. Experiments demonstrate that our method, Graph-RN with Adaptive Reference System Rotation (ARSR), achieves higher accuracy than baseline methods across multiple datasets (NTU-S, NTU-T, PBB, Taichi). In 5-way-1-shot and 5-way-5-shot tasks, it achieves competitive accuracy, such as 76.3±0.8% on PBB with 2s-AGCN (74.6±0.5% with ST-GCN). Attention heatmaps confirm the model’s focus on martial arts-relevant joints. The method provides an efficient solution for limited-sample action recognition, serving as a scientific martial arts training tool.
We propose a novel video inpainting algorithm composed of three tightly integrated components: a masked motion estimator, a recurrent hole filler, and a sequential video transformer. Unlike existing approaches that estimate motion from original frames, the proposed masked motion estimator predicts reliable bidirectional motion fields directly from masked frames, enabling principled motion estimation for video reconstruction. Based on the estimated motion, the recurrent hole filler propagates visual content and motion information over time while explicitly aligning completed and intact regions to reduce temporal inconsistencies. To further complete remaining holes, we introduce a sequential video transformer that inpaints frames recursively without partitioning a video into independent subsequences. The proposed transformer incorporates a unilateral attention mechanism tailored for sequential processing, achieving long-range temporal consistency with reduced memory and computational cost. Extensive experiments on standard video inpainting benchmarks demonstrate that the proposed method achieves state-of-the-art performance in terms of reconstruction fidelity and temporal consistency on long video sequences. The source code will be made publicly available.
With the advancement of Virtual Reality (VR) technology, Omnidirectional Images have gained widespread application due to their coverage of a spherical field of view and support for free viewpoint interaction. However, the Equirectangular Projection (ERP) introduces uneven distortion across different latitudinal regions. Concurrently, users typically navigate through local areas via continuous viewports, resulting in a pronounced spatial position dependency in the perceived quality of Omnidirectional Images. Traditional image quality assessment methods struggle to be directly applied to omnidirectional images, leaving Omnidirectional Image Quality Assessment (OIQA) facing considerable challenges. To address this, this paper proposes a no reference omnidirectional image quality assessment model based on viewport and latitude-aware distortion adaptive correction. The model adopts a dual-branch architecture: the local branch employs the XFeat keypoint detection method for adaptive viewport sampling of the Omnidirectional Image. A Vision Transformer network models spatial relationships between multiple viewports to capture local distortion features and their distribution information. The global branch directly processes ERP-formatted Omnidirectional Images, incorporating a Latitude-aware Distortion Adaptive Correction Module (LDACM) to effectively mitigate non-uniform distortion arising from spherical projection. Experimental results on OIQA and CVIQ datasets demonstrate superior performance of our proposed model compared to existing fully-referenced and no-reference approaches.
Low-light image enhancement frequently suffers from insufficient dark-region enhancement, over-exposure in bright areas, detail loss, and inefficient information propagation. Traditional Mamba-based models exhibit three main limitations that restrict low-light enhancement performance: luminance insensitivity, fixed scanning strategies, and lack of adaptive frequency-domain modeling, which restrict their performance in real-world low-light scenarios. To address these issues, this paper proposes a novel low-light image enhancement network, termed ISD-Mamba, which integrates illumination guidance, degradation-contrastive scanning, and frequency-selective modeling. The network follows a U-shaped encoder–decoder architecture with three core customized components: (1) Illumination-Guided State Space Modulation (IG-SSM) that employs multi-scale brightness priors to realize spatially adaptive feature modulation; (2) Degradation-Contrastive Scan (DC-Scan) that sorts spatial tokens by degradation degree and enables high-quality regions to guide the recovery of degraded regions; (3) Frequency-Selective State Propagation (FSSP) that adaptively decomposes features into low-frequency structures and high-frequency details with asymmetric state-space dimensions for balanced modeling. We further design an Innovative Mamba Block (IMB) to unify the above mechanisms and introduce illumination-aware skip connections for multi-scale feature fusion. Extensive experiments on LOLv1, LOLv2-real, and LOLv2-syn datasets show that ISD-Mamba obtains competitive performance with favorable balance between effectiveness and efficiency. The model has only 3.16 M parameters and runs at 0.042 s per 128 × 128 image, showing practical potential for real-world visual applications.
Medical images tend to exhibit blurred anatomical boundaries and similar intensity distributions across different organs or tissues, which reflects a high degree of self-similarity. Such characteristics hinder the effectiveness of current token reduction techniques applied to transformer-based medical image segmentation architectures, thereby limiting their applicability in point-of-care scenarios. To address this issue, we decouple self-similarity into intra-class and inter-class components and introduce a progressive token merging (PTM) method for medical Transformers. The key challenge is to (1) merge visually similar and semantically identical (i.e., intra-class) tokens to improve throughput, while (2) avoiding the merging of visually similar but semantically different (i.e., inter-class) tokens to maintain segmentation performance. Specifically, we perform token merging in two stages: visual token grouping (VTG) and semantic token matching (STM). At the initial layer, VTG merges tokens with similar appearance cues within a local window by average pooling, which suppresses noisy features and reduces the risk of incorrect inter-class merging. At the intermediate layer, STM leverages semantic cues in the tokens to construct a global bipartite graph and further reduces background or redundant tokens through bipartite matching. Through these two rounds of merging, PTM promotes a more balanced distribution of attention across semantic categories. Experimental results on two widely used medical image segmentation benchmarks demonstrate that the proposed method improves both segmentation quality and computational efficiency. The code and pretrained models will be publicly available at https://github.com/wuwen1994/PTM.
Defect detection in automotive bushings is important for suspension reliability and safety, yet production-line automation remains difficult. Vision-language models (VLMs) offer semantic interpretation beyond conventional pattern matching, but their memory footprint and power demand limit edge deployment. We propose a teacher-student knowledge distillation framework that transfers a large VLM’s semantics into a compact edge model. A Qwen3-VL-4B teacher receives dual-stream input pairing grayscale appearance with stratified height encoding from 16-bit TIFF surface data, with the student trained on response, feature, and attention objectives. Symmetric 8-bit integer (INT8) quantization compresses the student to 480 MB for NVIDIA Jetson Orin Nano. On 25,000 samples, the student reaches 91.3% mean average precision at 0.5 intersection-over-union (mAP@50), 94 ms latency, and 17.5 W power, retaining 93.4% of the teacher’s mAP@50 and exceeding YOLOv12 by 4.6 percentage points under the 100 ms real-time constraint, indicating that VLM-level semantic defect detection can operate on edge hardware.
To address the prevalent challenges of mode collapse and the fidelity-diversity trade-off in single-image generation, we introduce GSID, a multi-scale framework for structurally-guided synthesis via diffusion models. GSID utilizes a modified ConvNeXt Block to enhance structural regularization, effectively capturing global context while preventing overfitting in a data-scarce environment. We also introduce a High-Frequency Extractor Module (H-FEM) to improve the preservation of fine-grained details, achieving explicit structure-texture decoupling. Furthermore, we formulate the Dynamic Mask Strategy as a reactive closed-loop inference mechanism. Instead of relying on manually predefined static perturbation parameters, GSID periodically measures perceptual deviation during sampling, computes an error signal against a dynamic fidelity threshold, and adaptively adjusts the mask strength to regulate the generation trajectory. This feedback-regulated inference process enhances image diversity and controls structural variation without requiring additional training or per-instance manual parameter tuning. Evaluated on various natural and medical image datasets, GSID generates diverse, high-quality samples, demonstrating superior structural consistency and perceptual quality compared to existing state-of-the-art methods.
Recently, a marginal area that compensates for inaccurate viewport prediction is encouraged to compete for network resources, resulting in improved quality of experience (QoE) of 360-degree video. However, 360-degree video may suffer from a degradation in quality if the marginal area abuse network resources. To address this issue, we propose a multi-priority multi-path transmission framework (M2PT) for delay-constrained 360-degree video streaming in heterogeneous time-varying wireless networks. The proposed framework analyzes the importance of multi-priorities under dynamic network conditions to solve the trade-off between load balancing of different links and quality allocation of different viewports. Besides, a semi-physical simulation platform is established to verify the performance of M2PT. The experiment results show that the proposed M2PT provides better performances: a lower ratio of overdue frames, a reduction of the freeze ratio, and a higher average QoE quality, up to 3–6 dB for different video sequences compared to several reference schemes.
