Object detection in UAV imagery presents a significant challenge due to low image resolution, complex visual backgrounds, and substantial inter-object scale variation. Detection Transformer (DETR)-based models often underperform in such scenarios. This paper proposes Pre-Fusion Guided DETR (PF-DETR), a refined and enhanced extension of the D-FINE model, specifically designed for robust small-object detection in UAV imagery. PF-DETR introduces a novel Pre-Encoded Feature Fusion (PEFF) module, which employs a unidirectional fusion strategy to integrate low-level spatial features into high-level semantic features, thereby improving the model’s ability to capture fine-grained details of small objects. Additionally, a Feature Enhancement Attention (FEA) mechanism is incorporated to strengthen feature representation and reduce background interference. To address information loss during multi-scale processing, a Haar Wavelet Downsampling (HWD) module is proposed, combining Haar wavelet transforms with convolutional downsampling to better preserve critical features. Furthermore, the original GIoU loss is replaced with an Enhanced IoU (EIoU) loss function to improve localization accuracy and training efficiency. Experimental results on the challenging VisDrone2019 benchmark demonstrate that PF-DETR outperforms state-of-the-art detectors, including DEIM, D-FINE, YOLOv10, and YOLOv11, in both detection accuracy and robustness. Specifically, PF-DETR achieves an AP50 of 40.6
The Knowledge Graph (KG) augmented recommendation mitigates the cold start issue by exploiting complex semantic clues in users' behaviors. However, existing methods focus on bringing these auxiliary knowledge into the user-item interaction space, ignoring the gap between different sources. In this study, we propose a knowledge-aware intent subgraph learning (KISL) which mines users' intents with KG to promote fine-grained interest learning for personalized recommendation. We model each intent representation as an attentive combination of KG relations on the knowledge graph. Guided by the intents, KISL devises a dimensional disentanglement to divide the interaction graph into several augmented intent-aware subgraphs. Fine-grained personalized embedding is learned during subgraph message propagation to predict users' interactions. Extensive experiments on two public datasets demonstrate the effectiveness of KISL by knowledge-aware intent modeling over baselines.
Point-cloud denoising should suppress noise without harming geometry, classical approaches often oversmooth complex object surfaces, erasing fine details. Although recent learning-based techniques alleviate these shortcomings, they still face two critical limitations: (i) single-scale feature extractors without explicit positional encodings struggle to capture long-range dependencies, and (ii) stage-wise pipelines that update normals and coordinates separately accumulate errors. We propose an end-to-end framework that fuses multi-scale convolutions with a position-encoded Transformer and a geometry-guided feedback loop. Convolutions capture local detail, the Transformer models global relations, and residual fusion unifies their features. A key-point selector, driven by normal orthogonality and cross-scale agreement, retains only high-confidence points, while a bidirectional module jointly refines coordinates and normals under a geometric-consistency loss. On synthetic CAD and non-CAD datasets corrupted with 0.6– 2.0 % Gaussian noise, the method achieves lower Chamfer and point-to-surface distances than five state-of-the-art baselines. Ablations confirm each component’s value, and qualitative results preserve sharp edges and high curvature, providing an efficient, practical solution for point-cloud denoising.
Medical image segmentation task plays an important role in areas such as clinical care and medical atlas construction, but existing 2D segmentation models often face challenges in 3D organ or tissue segmentation tasks, and 3D models account for a large computing resource proportion. Therefore, we consider 3D tissue organs as video sequences and use the given first layer of tissue organ boundaries as a sequence of query points to predict matching points on subsequent slices. Based on this idea, we propose a point tracking architecture optimised based on the joint point tracking model CoTracker, which compensates for the shortcomings of CNN-based architectures in global feature information extraction and improves the robustness of the overall image features by incorporating the robust, strongly generalisable DINOv2 encoder. The model uses semantic features extracted by the self-supervised learning foundational model DINOv2-ViT for feature fusion with ResNet, and optimises them with a fine-tuning strategy based on the CoTracker weights. Specifically, we introduce the channel attention mechanism to make full use of Vision Transformer's feature recognition capability, and achieve feature optimisation by filtering high-weighted channels, thus improving the accuracy of several evaluation metrics in the point tracking domain. Extensive evaluation results show that our approach not only greatly saves training resources, but also efficiently improves tracking accuracy and has good generalisation in the field of medical image segmentation.
Single-stage 3D object detection based on pillar structures has gained attention for its high inference efficiency in autonomous driving. However, the quantization of point clouds into pillars often leads to the loss of fine-grained structural details, limiting performance in detecting small objects under sparse scenarios. To overcome this challenge, we propose an efficient LiDAR-based detection algorithm that incorporates multi-source feature encoding and structure-aware enhancement. Specifically, we design a Structure-Aware Feature Encoding (SAFE) module that integrates intra-pillar attention, inter-pillar structural modeling, and height histogram encoding to improve the geometric representation of pillar features. The backbone further incorporates a large kernel attention mechanism to capture long-range dependencies, along with an atrous spatial pyramid pooling module and a weighted dual-scale feature fusion module to strengthen semantic expressiveness and detection accuracy. The proposed method is evaluated on the KITTI dataset, with extensive visualizations and ablation studies. Results show an improvement of 3.9
The new generation video coding standard, H.266/VVC, introduces the quad-tree nested multi-type (QTMT) block partitioning structure and multiple inter coding modes, significantly improving coding efficiency but also increasing encoding time. To address this, we propose a coarse-to-fine fast partition decision (FPD) algorithm that collects both temporal and spatial information for inter CU partitioning. FPD first leverages co-located similarity between the current CU and its counterpart in the reference frame to capture global motion. High similarity indicates static regions, allowing early pruning of partition candidates. For CUs with low similarity, indicating complex local motions, we introduce a machine learning-based approach. Specifically, we extract temporal optical flow and spatial features (e.g., edges and gradients) to train a LightGBM classifier to predict the partition direction and skip the horizontal/vertical directions in advance. Experiments conducted under the common test condition of H.266/VVC demonstrate that our proposed FPD achieves a 37
Early diagnosis of Branch Atheromatous Disease is critical for reducing associated disability rates. Recent advancements underscore the potential of deep learning methods in developing automated diagnostic tools. However, the limited availability of clinical data poses a significant challenge to their practical application. Traditional deep learning models, which rely on large-scale datasets, perform poorly in few-shot scenarios, limiting their effectiveness in this context. To address these challenges, we propose a few-shot learning framework for Branch Atheromatous Disease diagnosis that leverages Lesion Localization prior knowledge. This approach incorporates domain expertise to guide data augmentation, effectively mitigating the issue of limited training data. Specifically, our framework is developed based on 251 BAD slices, 51 Non-BAD slices, and 400 lesion-free slices obtained from preprocessed clinical DWI images. Furthermore, we introduce a two-stage training strategy and an adapter module for parameter-efficient fine-tuning, enabling effective model optimization even with constrained data. Our method was evaluated on clinical cases from multiple medical centers, demonstrating superior diagnostic accuracy and robustness compared to various baseline models. These results highlight the potential of our approach to enhance the efficiency and accuracy of early BAD diagnosis and alleviate clinical diagnostic workload.
Graph Convolutional Network have emerged as a pivotal method in skeleton-based action recognition, demonstrating exceptional performance across multiple benchmarks. However, the inherent measurement errors in skeleton data-particularly the joint position estimation errors caused by occlusion-severely limit the recognition accuracy of existing models. To address this issue, this paper proposes a multimodal contrastive learning framework, PoseCLR, which effectively mitigates the impact of joint errors by leveraging the complementary characteristics of 2D and 3D skeleton data. Specifically, 3D skeleton data provides rich spatial information in three dimensions, while 2D skeleton data preserves more precise two-dimensional joint coordinates. The feature interaction between these two modalities achieves error compensation and information enhancement. Furthermore, this work supplements complete information on joint motion and bone motion, significantly improving the feature representation capability of individual modalities and thereby enhancing the performance of multi-stream score fusion. Experimental results demonstrate that the proposed method achieves breakthrough performance improvements on three mainstream datasets-NTU RGB+D, NTU RGB+D 120, and NW-UCLA-reaching current state-of-the-art methods.
Current cross-modal retrieval models predominantly rely on one-to-one image-text pairs for training, with most approaches projecting each sample into a single embedding vector. However, recent studies indicate that this single-vector representation fails to capture the inherent complexity of images and texts, resulting in the loss of critical semantic information during the embedding process. Consequently, these models often struggle to effectively learn the nuanced features of both modalities. In this paper, we propose a novel training strategy that transitions from traditional one-to-one image-text pairing to a one-to-many framework and introduce an innovative Set Prediction Module with Weight to better capture the diverse semantics of the input data. By incorporating a more diverse set of textual representations during training, our method significantly enhances the performance of cross-modal retrieval models. Extensive evaluations across various backbone networks on the COCO and Flickr30K datasets demonstrate that our approach consistently outperforms most existing methods.
Current air combat decision methods mostly lack structured adversarial validation and adaptive reflection mechanisms, limiting their robustness and reliability when facing deceptive or unfamiliar enemy strategies. To address this, we propose the Adversarial Iterative Pre-Enactment (AIP) framework, which integrates cognitive mental simulation with a large language model. First, we design a self-adversarial pre-enactment–feedback loop that enables the model to simulate and evaluate both friendly and enemy actions before execution. Second, we introduce a multiscale segmented evaluation mechanism that analyzes tactical effectiveness across different time horizons and perspectives. Third, we build a high-fidelity simulation environment with scenario rewind capability, enabling real-time dual-strategy execution and supporting structured adversarial evaluation. Experimental results demonstrate the superiority of AIP over standard LLM baselines in terms of both tactical score and stability under complex adversarial conditions.
Commonsense video captioning requires the model not only to describe visible content but also to infer multiple types of commonsense captions, including "Intention", "Effect", and "Attribute". Existing methods generally fall into two categories: one extracts commonsense directly from videos but struggles to bridge the semantic gap between visual content and implicit commonsense under limited knowledge; the other leverages language model-extended textual knowledge, which alleviates this gap but overlooks the semantic relationships among different types of commonsense information, limiting reasoning capability. To address these challenges, we propose a Semantic Relation-Guided Network (SRG-Net) for commonsense video captioning. Specifically, a Commonsense Semantic Relation Modeling (CSRM) module is designed to capture interrelations among different types of extended commonsense knowledge and enhance their representations. Furthermore, a Hierarchical Fusion Decoding (HFD) strategy is adopted. Multimodal video features are first fused, followed by the integration of enhanced commonsense representations, enabling the generation of accurate and fluent commonsense captions. Extensive experiments on the large-scale Videoto-Commonsense dataset demonstrate that SRG-Net achieves superior performance compared to existing methods across multiple metrics.
Recently, several newest Depth Map Super-Resolution (DMSR) methods have incorporated depth edge prediction as auxiliary guidance into the optimization model, creating a dual-task driven unfolding network for enhancing depth edge refinement. Nevertheless, these approaches either overlook explicit dual-edge consistency constraint or merely fuse color edges with depth edges once. This oversight results in diminished generalization ability and subpar performance. To this end, we transform the DMSR problem as a triple-task optimization model explicitly constrained by dual-edge consistency. According to the Alternating Direction Method of Multipliers (ADMM) theory, the proposed model can be cast as iterative sub-optimizations for color-edge update, depth-edge update, depth map update, and augmented Lagrange multiplier update. These sub-optimizations can be further unfolded into an interpretable ADMM network. Within this network, we integrate learnable modules into the initial pure formula expansion, enabling high-throughput information transmission and thereby enhancing the network's representational power. A large number of experiments have demonstrated that the proposed method achieves better reconstruction results as compared with several DMSR methods.
To address the accuracy-speed trade-off in semantic segmentation for resource-constrained driver assistance systems, this study proposes EAANet (Edge-Aware Attention Network). Building upon a Lightweight Progressive Scalable Network (LPSNet) as the baseline, we first embed Squeeze-and-Excitation (SE) channel attention into feature fusion layers to establish channel-wise adaptive feature selection. The Atrous Spatial Pyramid Pooling (ASPP) module is then enhanced through four parallel branches for improved multi-scale feature extraction. Further optimizations include reconstructing the decoder with depthwise separable convolutions and compressing the model size via parameter pruning. Experiments on the Cityscapes dataset demonstrate that EAANet achieves 76.7
Glioblastoma (GBM) is the most aggressive glioma with a 5-year survival rate of only 6.8
Due to the strong scattering and low absorption of light in biological tissues, the inverse problem of bioluminescence tomography (BLT) is ill-posed. Hybrid regularization constraints can effectively alleviate the inherent ill-posedness of the BLT inverse problem. However, the hybrid regularization of traditional algorithms involves multiple parameter determinations, and it is difficult to select parameters manually. This paper proposes deep unfolding based on U-Net graph-guided hybrid regularization (DUnet-GGHR) method for BLT. The gradient update step in the traditional GGHR algorithm is reformulated as a flexible gradient descent update module for automatic learning. At the same time, the soft threshold calculation step in the traditional GGHR algorithm solution process is expanded into a proximal mapping module for training. A series of reconstruction experiments have verified that the proposed DUnet-GGHR model performs better than other end-to-end deep unrolling comparison methods in tumor localization, morphology restoration, and energy recovery. At the same time, the mathematical process is combined with the deep neural network to improve the interpretability and generalization of the network, reduce training data, and speed up the calculation.
The rapid development of industries such as heavy machinery, shipbuilding, and energy equipment has led to a sustained increase in the demand for medium-thick plate welding. However, conventional welding methods still heavily rely on manual operations or robotic teaching, resulting in low efficiency and poor consistency in weld quality. These limitations make it difficult to meet the high precision and productivity requirements of modern intelligent manufacturing. To address these challenges, this study proposes an intelligent multi-layer, multi-pass welding system for medium-thick plates, integrating deep learning and adaptive control. First, a deep convolutional neural network (CNN) based on the ResNet101 architecture was developed to automatically classify weld groove types, achieving a classification accuracy of 99.62
Traditional visual synthesis methods suffer from blurring, ghosting, and significant artistic deprivation due to their sole reliance on physical disparity reconstruction. These methods fail to capture the artistic elements that directors carefully design in classic 3D films. To address these issues, we propose Art3D-Fusion, a hybrid framework combining geometric processing with advanced diffusion models. This framework integrates physical depth information with global artistic control, particularly in 0-plane selection. Through depth estimation and optical flow-based disparity matching, we create a pseudo-realistic disparity map that reflects the director's artistic adjustments. This approach enables precise depth-to-disparity transformation and ensures natural detail restoration and effective occlusion handling. Art3D-Fusion generates high-quality right-view images by driving a diffusion model under an enhanced ControlNet architecture using the artistic disparity map and original left-view image as dual-conditional inputs. This approach accurately represents the director's artistic vision and allows for the transfer of artistic styles from classic 3D films to new scenes, ensuring the consistent reproduction of different directors' artistic styles through the estimation and transfer of camera parameters.
Magnetic Resonance Imaging (MRI) plays a vital role in medical and biological applications. However, for patients with MRI-compatible implantable devices such as cochlear implants, the presence of integrated magnets often leads to large signal voids and severe artifacts, significantly compromising diagnostic accuracy. Although recent advances in deep learning have shown promise in artifact reduction and image enhancement, the quantitative assessment of artifact regions still heavily relies on manual annotation, which is labor-intensive and inconsistent. In particular, boundary distortions and tissue loss near cranial regions pose significant challenges for accurate artifact delineation, limiting the effectiveness of existing segmentation methods. To address these issues, we propose a novel 3D artifact segmentation framework that integrates reflective registration into a deep neural network combining U-Net and Transformer architectures. We conducted experiments on MRI data from 5 real-world patients with cochlear implants. Experimental results demonstrate that our method achieves state-of-the-art performance in implant-induced artifact segmentation, offering an efficient and reliable solution for automatic artifact evaluation in clinical settings.
Cardiac Magnetic Resonance Imaging (CMR) is a crucial clinical imaging modality for assessing cardiac morphology and function, and it has become the gold standard for diagnosing cardiovascular diseases. However, its widespread clinical adoption is hindered by long acquisition times and high costs-challenges that are particularly acute in dynamic imaging, where both high spatial and temporal resolutions must be achieved within a limited timeframe. In this paper, we propose a novel dynamic deep unrolling method, CMambaR, a cardiac phase-embedded Vision Mamba architecture designed to accelerate cardiac MRI reconstruction. The proposed method integrates the strengths of unfolded iterative optimization with a spatiotemporal dynamic reconstruction network, enabling it to effectively capture complementary information embedded in dynamic sequences while leveraging physics-based priors to deliver high-quality reconstruction. Inspired by structured state space models, we design a local enhanced vision Mamba module as the core building block of our network, capable of capturing both local details and long-range dependencies. Furthermore, we introduce a cardiac phase fusion mechanism that incorporates cardiac phase prior into the reconstruction process, further enhancing reconstruction performance. Extensive experiments on two cardiac datasets demonstrate that our method achieves high-fidelity image reconstruction and consistently outperforms existing approaches.
Microvascular decompression (MVD) plays a critical role in the treatment of neurovascular compression-related diseases, with its success heavily dependent on the precise preoperative identification of key anatomical structures, especially small-volume and densely distributed tissues like nerves and vessels. To address this challenge, we propose a multi-attention aggregation network (MAA-Net), for the segmentation of MVD-related structures in MRI images. The method is based on the U-Net architecture and incorporates two attention mechanisms. A spatial-channel parallel attention module at the bottleneck jointly models spatial and channel dependencies to better capture complex interwoven anatomical structures, particularly in regions where nerves and cerebral vessels are intertwined. In addition, a lightweight gated attention module is inserted between the encoder and decoder to improve the perception of small-volume nerve targets by promoting feature selectivity and suppressing background noise, especially when processing fine nerve structures. We evaluated the proposed method on a private clinical dataset covering the brainstem, nerves, cerebral vessels, and cerebellum. The results demonstrate that our method achieves superior performance in volume overlap metrics (e.g., Dice score), delivers precise boundary delineation in distance-based metrics (HD95 and ASD), and exhibits strong recognition ability for elongated structures as reflected in the clDice score, confirming its practical value for preoperative MVD localization.
