Accurate segmentation of thyroid nodules in ultrasound images is essential for thyroid cancer risk assessment and computer-aided diagnosis, yet remains challenging due to ambiguous boundaries and significant shape variations. Existing convolutional neural network (CNN)-based methods effectively capture local features but are limited in modeling long-range dependencies and complex boundary structures. To address these limitations, we propose SAFM-Net, a dual-branch CNN–GNN Network that integrates synergistic attention and frequency-domain modulation. The network adopts a dual-branch encoder, where a graph-based branch leverages a synergistic-attention dynamic graph convolution (SA-DGC) module to adaptively model global relationships among feature nodes, enhancing structural and boundary representation. In parallel, a CNN branch captures local textures and fine-grained details. To fuse complementary features, a frequency-domain modulation (FDM) module is introduced to enable cross-branch interaction and hierarchical integration, improving feature representation capability. Extensive experiments on the DDTI and TN3K datasets demonstrate the effectiveness of the proposed method. Compared with GED-Net, SAFM-Net achieves improvements of 0.54%, 1.00%, 1.68%, and 0.73% in terms of Accuracy, Dice, IoU, and Precision, respectively, on the DDTI dataset, and improvements of 0.17%, 0.43%, 0.67%, and 1.48% in terms of Accuracy, Dice, IoU, and Precision, respectively, on the TN3K dataset. These results indicate that SAFM-Net provides accurate and robust segmentation performance under challenging ultrasound imaging conditions.
Object detection in low-light environments is severely degraded by photon starvation, leading to feature concealment, noise amplification during feature fusion, and motion blur caused by long exposure. To address the limitations of the conventional “enhancement-then-detection” pipeline, this paper proposes CIE-Det, an end-to-end lightweight object detection framework that jointly optimizes illumination enhancement and feature learning. A lightweight cascaded illumination enhancement module (CIE module) with 0.0017 M parameters is first embedded at the network input to restore image contrast via downsampled illumination estimation and cascaded nonlinear curves with a dynamic gating mechanism. To alleviate cross-layer semantic inconsistency, a dynamic semantic alignment fusion operator is introduced to replace static feature concatenation with learnable adaptive weighting, enabling effective noise suppression and multiscale feature alignment. Furthermore, an orthogonal asymmetric large-kernel detection head (OAK-Detect) is designed using separable 1×5 and 5×1 convolutions to enhance receptive fields and improve robustness to motion blur. Extensive experiments on the ExDark dataset demonstrate that CIE-Det achieves 73.1% mAP@0.5 (mean average precision at an intersection over union threshold of 0.5), outperforming YOLOv11s by 2.5%, while reducing model complexity to 9.12 M parameters. The proposed method achieves a favorable trade-off between detection accuracy and efficiency in low-light scenarios.
Handwritten mathematical expression recognition (HMER) aims to convert images of handwritten mathematical expressions into structured LaTeX sequences. Although recent encoder-decoder models have achieved strong performance, they still struggle with visually similar symbols, Greek letters, special symbols, and structural tokens such as fractions, radicals, superscripts, and subscripts, especially when expressions contain complex spatial layouts. To address these issues, we propose a Counting-based Position-aware Transformer (CP-Former) that integrates symbol-counting supervision and position-aware structural modeling into a CoMER-based recognition framework. Instead of treating counting and position information as isolated cues, CP-Former jointly learns global symbol occurrence statistics and relative structural positions from LaTeX annotations, without requiring additional symbol-level labels. Experiments on CROHME 2014, 2016, and 2019 show that CP-Former consistently improves over the CoMER baseline, with absolute ExpRate gains of 3.00%, 0.96%, and 2.00%, respectively. Compared with recent state-of-the-art methods such as PosFormer, CP-Former achieves competitive exact-match performance and obtains comparable or better results on several error-tolerant metrics, particularly under the ≤2 and ≤3 settings. These results suggest that the proposed counting-based position-aware modeling improves structural robustness for complex handwritten expressions, whereas the fair ablation study further shows that the gains cannot be attributed to counting supervision alone, but mainly arise from its integration with position-aware structural modeling.
Due to complex underwater illumination and scattering effects, underwater images commonly suffer from color distortion, low contrast, and structural blur. These distortions not only reduce visual perceptual quality but also severely limit the performance of downstream object detection tasks. Although existing underwater image enhancement methods have made notable progress in improving human visual perception, their optimization objectives mainly focus on subjective visual quality while overlooking detection-friendly feature representations. This results in a distributional mismatch with detector features, ultimately degrading object detection performance. To address these issues, we propose a global-local feature collaborative underwater image enhancement network. The network adopts a three-branch collaborative architecture to model local structural details, cross-color channel dependencies, and global contextual information, enabling joint optimization of global consistency and local discriminability. It provides more stable and discriminative feature representations for downstream object detection. To enhance the perception of small-scale targets and fine-grained structures, we design a structural detail enhancement module (SDEM) that combines multibranch and dilated convolutions to capture edge and texture features at different scales. The detail recalibration attention module (DRAM) combines local feature refinement with channel-spatial attention to adaptively reweight multibranch features, enabling fine-grained fusion of heterogeneous enhanced features and reinforcing key structural cues in the enhanced images. The experimental results show that our method significantly improves detection accuracy on the DUO and URPC2020 datasets. It also achieves strong PSNR and SSIM performance on the UIEB dataset. These results demonstrate that the proposed method can balance the detection performance and the visual quality.
Underwater three-dimensional reconstruction is inherently challenged by light scattering and wavelength-dependent absorption, causing color distortion, contrast degradation, and limited visibility in recovered geometries. Existing underwater three-dimensional (3D) Gaussian splatting methods rely solely on pixel-wise losses [L1/structural similarity index measure (SSIM)], which produce over-smoothed textures in scattering environments, while their direction-only medium modeling assumes spatially homogeneous water properties that contradict real-world depth-dependent turbidity. To address these limitations, we present an enhanced framework for underwater 3D reconstruction integrating progressive perceptual supervision with spatial coordinate encoding. A progressively weighted learned perceptual image patch similarity (LPIPS) loss recovers fine texture details without compromising geometric accuracy, while sinusoidal pixel-coordinate encoding enables self-supervised modeling of spatially varying turbidity without depth sensors. Experimental evaluation on the SeaThru-NeRF dataset demonstrates substantial perceptual quality improvements, with the progressive LPIPS alone achieving up to 33.9% LPIPS reduction on high-turbidity scenes, while the full framework maintains photometric accuracy (peak signal-to-noise ratio/SSIM). Cross-dataset evaluation on the out-of-distribution, deep-sea Submerged3D dataset further confirms that these perceptual gains generalize beyond a single benchmark. The method achieves interactive rendering performance (15 to 35 frames per second) and efficient training (∼8 to 13 min), providing a practical solution for real-time underwater 3D reconstruction in robotic applications.
Occlusions and far distance pedestrians pose significant challenges for pedestrian detection, often leading to insufficient feature representation learned by models, which in turn results in degraded detection accuracy and a high miss rate. To address this issue, we propose a three-stage fusion multispectral pedestrian detection network named TFNet. The network employs a three-stage fusion strategy: first, the multiscale feature fusion attention module selectively enhances critical features within each modality, effectively focusing on discriminative regions of occluded and distant pedestrians. Subsequently, the transformer interactive fusion module establishes long-range dependencies and enables dynamic interaction across modalities, achieving deep semantic alignment and complementary information exchange. Finally, the pixel-adaptive feature fusion (PAFF) module performs pixel-level refinement and adaptive fusion of the interacted features, generating a more discriminative unified representation. Extensive experiments conducted on the public multispectral pedestrian detection KAIST dataset show that TFNet outperforms existing state-of-the-art algorithms. In particular, it achieves a steady and substantial reduction in miss rate on the occlusion subset and multiscale subset. Meanwhile, it obtains higher average precision on the FLIR and LLVIP datasets. All experimental results fully demonstrate that the proposed network can effectively alleviate missed detections of occluded and far-distance pedestrians, and is expected to present high application value in the field of pedestrian safety monitoring in public scenarios.
The corpus callosum (CC), the largest white matter structure in the human brain, plays a vital role in interhemispheric communication and is closely associated with several neurological disorders. Accurate morphometric analysis of the CC can provide important insights into disease-related structural alterations. However, manual segmentation of the CC from MRI scans is time-consuming and subject to interobserver variability, limiting its applicability in large-scale neuroimaging studies. To address these challenges, we present DeepMorph-CC, an automated deep learning framework for segmentation and quantitative morphometric analysis of the CC from 2D T1-weighted MRI images. Rather than introducing a new segmentation backbone, the proposed approach focuses on the task-specific adaptation of an attention-guided U-Net architecture and its integration into a unified analysis pipeline. Spatial attention mechanisms are employed to improve boundary localization and segmentation accuracy, whereas K-fold cross-validation has been utilized to enhance model robustness under limited training data conditions. Following segmentation, anatomically consistent CC parcellation has been performed to enable region-wise morphometric analysis. Key structural features, including area, circularity, solidity, and major and minor axis lengths, are extracted for the entire CC and its subregions. The proposed framework is evaluated using two publicly available datasets, the ABIDE dataset and the OASIS dataset, along with a real clinical MRI dataset (RCI). Experimental results achieve an average DSC of 98.70 +/- 0.37%, demonstrating high segmentation accuracy. Consistent morphometric patterns across datasets and strong agreement with expert annotations validate the effectiveness and reliability of the proposed framework. Future work will focus on extending the framework to multicenter datasets and three-dimensional MRI volumes
Gland segmentation is pivotal for pathological diagnosis yet remains hindered by morphological heterogeneity and instance adhesion. We present amplitude-phase corrective and edge-refined network (ACE-Net), a multi-domain synergistic framework leveraging frequency-domain correction and edge guidance for high-precision segmentation. To bolster feature representation and context modeling, we develop parallel convolutional blocks and pyramid-aware context modules inspired by nonlinear spiking neural P (NSNP) theory, enhancing structural aggregation and nonlinear characterization. To mitigate spatial frequency decoupling, a phase-amplitude interactive module is integrated with deep-supervised edge heads to enforce spectral consistency constraints. Furthermore, a DGF decoder driven by learnable directional priors, in conjunction with a dual-layer edge-auxiliary supervision mechanism, facilitates meticulous boundary reconstruction through cross-scale gradient guidance. Empirical results on GlaS and CRAG benchmarks indicate that ACE-Net consistently outperforms state-of-the-art architectures in segmentation accuracy, boundary fidelity, and structural consistency. We extend the NSNP-based multi-domain modeling paradigm, offering an efficient and interpretable solution for complex histological structures.
Point cloud denoising is a fundamental prerequisite for downstream 3D perception and reconstruction. Existing methods often improve either global smoothness or local detail fidelity, but struggle to preserve both under heavy noise. We propose a dual-stage gated point cloud network (DGPCN), in which denoising is explicitly decomposed into coarse structural recovery and fine residual correction. The first stage employs multiscale dynamic graph convolution together with a residual decoder to recover the global geometric topology, whereas the second stage introduces a dual-threshold spiking graph refinement branch to iteratively accumulate and correct subtle residual deviations in high-curvature regions. A point-wise adaptive gating mechanism then fuses the coarse and refined displacements, enabling different points to rely on different stages according to their local geometric requirements. The temporal dimension is interpreted as an iterative refinement depth rather than physical time, and the dual-threshold neuron dynamics are explicitly formulated. Experiments on synthetic (PCNet and PU-Net) and real Kinect datasets, together with time-step sensitivity, multiscale neighborhood analysis, and extended ablation studies, demonstrate that DGPCN achieves consistently stronger denoising performance and better detail preservation than recent competing methods.
Images captured in low-light conditions suffer from severe detail loss, noise, and underexposure, significantly degrading downstream vision tasks. While supervised low-light image enhancement (LLIE) methods achieve strong performance, their reliance on scarce paired data limits practical deployment. Unsupervised approaches alleviate this issue but are typically confined to the red, green, blue (RGB) domain, lacking robust structural priors, which leads to uneven illumination and blurred boundaries. To address these limitations, we propose E2F-Net, a novel edge-to-feature paradigm for unsupervised LLIE that leverages depth-derived structural priors. Specifically, we introduce a depth-edge prior module (DEPM) that estimates scene depth and extracts depth-induced edge priors, injecting illumination-invariant geometric boundaries into the enhancement pipeline at an early stage. Building upon this structural foundation, a cross-modal fusion module (CMFM) employs a depth-guided attention mechanism, where depth features serve as queries to selectively aggregate complementary appearance and geometric information, enabling spatially consistent feature representations. Finally, a depth-aware enhancement module (DAEM) performs depth-adaptive residual refinement through a dual-gating strategy that dynamically balances geometric priors and visual features, facilitating robust illumination correction, structural detail preservation, and noise suppression. Extensive experiments on standard low-light benchmarks demonstrate that E2F-Net consistently outperforms recent state-of-the-art unsupervised LLIE methods in both visual quality and structural fidelity.
Remote sensing hyperspectral imagery is often degraded by complex mixed noise including gaussian noise, sparse impulse noise, structured noise such as stripes and deadlines. These degradations severely affect spectral signatures and spatial consistency, limiting the reliability of downstream tasks such as land cover mapping, military applications, environmental monitoring, and target detection. We propose HySTaR-Net, a two-stage hybrid denoising framework designed for hyperspectral remote sensing imagery under heterogeneous mixed noise conditions. In the first stage, spectral covariance-based subspace projection exploits the intrinsic low-dimensional structure of hyperspectral data to suppress dense band-wise Gaussian noise. Nonlocal low-rank and sparse modeling within the subspace domain further removes structured and sparse noise, whereas a noise-aware, adaptively scheduled, directionally weighted three-dimensional TV refinement further suppresses residual fluctuations under controlled regularization. The second stage applies a lightweight adaptive edge-guided spectral-spatial attention network in a residual learning manner to refine the intermediate result by correcting residual noise and model-induced artifacts while preserving spectral priors. Experimental results on benchmark remote sensing hyperspectral datasets demonstrate that the proposed framework consistently improves reconstruction accuracy, structural preservation, and spectral fidelity under challenging mixed-noise scenarios.
Road-scene object detection is a key component of perception for autonomous driving. In complex traffic with large-scale variation, heavy occlusion, and strong background clutter, small-scale and shape-complex objects are prone to missed detections, false positives, and localization errors, while existing models often struggle to balance accuracy and complexity. To address these challenges, we propose lightweight small-object & multiscale feature-fusion YOLO (LSMF-YOLO), a multiscale feature fusion detection network for complex driving scenarios. We design a lightweight HGNetV2 backbone by incorporating Light_HGBlock and CBAM convolution, enabling effective feature extraction with reduced parameters. In the neck, we adopt an SDI-based feature fusion structure and add a small-object detection layer to improve the separability and detection rate of small-scale complex targets. In the head, we incorporate feature-assisted spatial feature fusion head into the detection head for cross-level feature interaction and adaptive multiscale fusion, improving classification reliability and reducing localization errors. Extensive benchmarks on KITTI and BDD100K demonstrate that LSMF-YOLO achieves strong accuracy and generalization. With only 2.92M parameters, it attains 92.9% and 44.7% mAP50 and 89.2% and 51.1% F1, respectively. Compared with YOLOv11n, it improves mAP50 by 4.6% and 2.1% and F1 by 3.7% and 8.8%, confirming efficient perception under resource constraints and a better accuracy-efficiency trade-off.
Low-altitude unmanned aerial vehicle (UAV) remote sensing constitutes a pivotal paradigm for ecological monitoring; however, the detection of small wildlife targets is persistently hampered by fine-grained feature dissipation and semantic aliasing inherent to dense distributions and complex aerial backgrounds. To address these bottlenecks, we propose OGCF-YOLO, a lightweight, high-precision object detection network tailored for UAV edge deployment. Using YOLOv8n as a baseline, the method achieves a deep synergy between detection accuracy and computational efficiency by reconstructing feature transmission links and optimizing fusion mechanisms. The primary improvements include: (1) structural reallocation of computational resources: The removal of the redundant deep P5 branch and the introduction of a high-resolution P2 detection head to physically suppress the premature loss of small target information; (2) structured gated fusion architecture: the design of the C2f-OGCF module to reconstruct the neck network, where OrthoMix performs numerically stabilized channel-pair mixing and the gated competitive mechanism alleviates semantic conflicts during cross-scale fusion; (3) feature purification mechanism: the integration of the spatial and channel synergistic attention module at the end of the backbone to enhance feature separability for dense targets in complex backgrounds; and (4) content-aware dynamic upsampling: the integration of the DySample module to replace traditional interpolation, significantly improving fine-grained texture recovery capabilities. Experiments on the Animal Object Detection dataset demonstrate that OGCF-YOLO achieves a performance leap while substantially reducing computational overhead. Compared with the baseline model, the parameter count is reduced by 69.9% (to 0.77M), whereas precision, recall, mAP50, and mAP50:95 are improved by 3.9%, 12.9%, 14.1%, and 8.5%, respectively, reaching 85.5%, 70.5%, 80.4%, and 42.6%. Crucially, the model attains a single-frame inference latency of 104 ms on an NVIDIA Jetson Nano via TensorRT INT8, validating its efficacy as a high-recall solution for real-time ecological monitoring on resource-constrained UAV endpoints.
Crop disease detection is crucial for improving crop yield. Current methods mainly focus on single crops, and limited research has addressed challenges such as high inter-crop disease feature similarity, large-scale variations, and low detection accuracy in real field environments. In addition, high local deployment costs remain an issue. To tackle these challenges, we propose a lightweight model, MSN-Mamba-YOLO, for multi-crop and multi-disease detection. First, in the backbone network, an efficient feature extraction module, RG-MMamba, is constructed based on MobileMamba. This module combines wavelet convolution, Mamba, and multi-kernel depthwise separable convolution, effectively alleviating problems caused by high feature similarity and large-scale variations among diseases. Second, to improve detection accuracy in complex field environments, a dual-path parallel scanning mechanism, SN-XSSBlock, is designed in the neck: one path performs global scanning, and the other performs self-structure-aware scanning. By fusing global and structural perception, it achieves precise localization of diseased regions. Finally, in the detection head, a YOLOv11 head integrated with multiple depthwise separable convolutions is adopted, enabling more efficient feature decoupling, reducing computational overhead, and improving overall detection efficiency. On the FieldPlant dataset enhanced with generative-adversarial-network-based augmentation, MSN-Mamba-YOLO achieves an mAP@50-95 of 55.4%, a 1.5% improvement over Mamba-YOLO-T, while reducing parameter count and computational cost by 22% and 30%, respectively. In addition, the model's cross-crop generalization ability was validated on tea and eggplant disease datasets, demonstrating its efficiency and practicality in multi-crop, multi-disease detection scenarios.
Copy-move forgery is a common image manipulation operation in which one or more regions are copied and pasted within the same image to conceal or emphasize specific content. Among existing copy-move forgery detection (CMFD) methods, point-based approaches are attractive because of their efficiency and flexibility; however, they still face two major limitations: conventional keypoint extraction often produces excessive points in textured areas but insufficient coverage in smooth regions, and the matching stage still lacks a sufficiently robust and efficient mechanism under complex attacks. To address these issues, we propose a robust CMFD framework based on dominant-rotated local binary patterns (DRLBP) and hierarchical multidistance matching. First, simple linear iterative clustering (SLIC) and robust mean-shift (RMS) clustering are jointly used to achieve region-aware and spatially balanced keypoint extraction. Second, accurate quaternion-generalized orthogonal Fourier-Mellin moments (AQGOFMMs) are adopted to provide robust and discriminative feature representation. Third, DRLBP-guided hierarchical multidistance matching is introduced to improve matching efficiency and reliability under geometric and post-processing distortions. Finally, mismatched correspondences are removed through region-specific thresholding, A-RANSAC-based local transformation estimation, and optimized mean-residual normalized intensity correlation (NNPROD) localization. Extensive experiments on multiple benchmark datasets demonstrate that the proposed method achieves strong robustness, accurate localization, and competitive overall detection performance compared with representative traditional and recent methods. These results indicate that the proposed framework provides an effective and interpretable solution for practical copy-move forgery detection.
In archeology, the reconstruction of ancient fragments (pottery, mosaics, and marble tablets) is often a real challenge for experts studying the subject. This is due to the natural degradation of the materials (water, wind, and friction) over hundreds or thousands of years. These pieces, originally inseparable, are now devoid of any visual logic of assembly. Certain visual and geometric clues enable the most obvious assemblies to be made, but this remains limited to human perception and requires more precise processing of geometric features. We propose a method for matching surfaces by mapping geometric features such as curvatures, peaks, and surface normals, extracted by segmenting fragments into simple primitives. This mapping method is not dependent on the rotation or translation of the fragments. We present the first results obtained on two datasets and conclude with the ongoing optimization process designed to handle the real archeological database.
The enhanced compression model (ECM), an advancement beyond the versatile video coding (VVC) standard, further improves coding efficiency but introduces additional complexity in encoding. Similar to VVC's quad-tree with nested multitype tree partitioning, ECM adopts a flexible and adaptive partition structure that enhances compression performance but at the expense of greater encoding time owing to the computationally intensive rate-distortion cost evaluations. To address this challenge, we propose an efficient intra-partitioning algorithm that leverages deep-learning (DL) techniques through a convolutional neural network (CNN). Specifically, the method implements two CNN models to predict horizontal and vertical binary tree partitions at a 32 & times; 32 coding unit, which are then combined into a unified fast partitioning decision process for the ECM. Experimental results demonstrate that, compared with the ECM-10.0 reference software, the proposed CNN-based approach achieves up to 45.20% encoding time savings, with only a minimal impact on compression performance.
To address the challenges of small vehicle targets, complex backgrounds, and insufficient utilization of complementary information between visible and infrared modalities in remote sensing images, we propose a multimodal remote sensing vehicle detection algorithm based on Detection Transformer (DETR). The algorithm employs MobileMamba as a lightweight backbone network to extract multiscale features, achieving linear-complexity global modeling through a 2D Selective Scan mechanism. A Cross-Modal Global Fusion module is designed to extract modal statistical information via global average pooling and global max pooling, enabling deep interaction between visible and infrared features through a cross-attention mechanism. A Spatial Channel Feature Aggregation module is introduced to enhance multiscale feature representation by synergistically combining spatial attention and coordinate attention. A Separable Dynamic Decoder is constructed, replacing standard cross-attention with dynamic convolution attention to accelerate query-feature interaction. Experimental results on the DroneVehicle dataset demonstrate that the proposed method achieves an average precision (AP) of 72.8%, surpassing the baseline model by 3.2 percentage points, with a small-object detection metric (AP(S)) of 51.2%, while maintaining a real-time inference speed of 56 FPS. Cross-dataset validation on the VisDrone2019 benchmark further confirms the generalization capability of the proposed method, achieving 49.2% mAP@0.5 and surpassing existing state-of-the-art methods, thereby validating the effectiveness and broad applicability of the proposed approach.
Hashing methods in image classification and retrieval convert high-dimensional feature vectors into compact codes, improving storage efficiency and search speed. Learnable hashing improves class separation by minimizing intraclass and maximizing interclass distances but faces challenges. Most approaches maintain similarity between feature vectors and hash codes through extensive training and complex loss functions. In transfer learning, generating distinctive hash codes can lead to model overfitting due to the limited number of target domain samples. To address this, a common approach is to use features from the fully connected layer as the hash unit. In hashing-based methods, the presence of irrelevant or unimportant neurons in the output layer can negatively affect similarity computation and the retrieval of relevant samples. Conversely, limiting the number of neurons in this layer reduces the complexity of the hash unit, resulting in a lightweight architecture that enables model training with fewer samples. However, excessive compression of hash codes due to aggressive neuron reduction may degrade model performance. We investigate the optimal selection of the number of neurons in the output layer of the hash unit by exploring the trade-off between hash code compression and retrieval performance using the Boruta-Shap algorithm. The results demonstrate that the Euclidean distance becomes more discriminative as the number of neurons in the hash unit is reduced, thereby improving the retrieval of images similar to the query. Experimental results show that reducing the number of neurons from 1920 to 698 (63% reduction) improves retrieval accuracy on CIFAR-10. Furthermore, quantitative analysis indicates that the average Euclidean distance between query images and retrieved samples is reduced by up to 44% (from 9.16 to 5.21), whereas the average cosine similarity increases by up to 8% (from 0.72 to 0.78) compared with using a fully sized hashing layer.
