Yuheng Li

Accurate learning-based solvers for coupled reaction–diffusion systems should preserve mathematical structure rather than merely reduce pointwise residuals. We develop a structure-preserving deep Galerkin method that combines weak residual minimization with positivity, mass-balance, and energy-dissipation constraints. Adaptive quadrature concentrates test points near fronts and structural defects, while staged optimization stabilizes stiff training. Under coercivity, one-sided reaction monotonicity, and entropy compatibility, we establish an energy stability estimate, a weak-residual a posteriori error bound, and a unified structural reliability inequality. Numerical studies on pattern formation, nonlinear diffusion, and heterogeneous media show lower reported aggregate solution errors and structural drift than conventional deep Galerkin, variational, and physics-informed neural solvers; no inferential significance claim is made because seed-resolved records are not available in the retained archive. The results support invariant-aware variational learning as a reliable framework for nonlinear PDE approximation.

Visible–infrared detection improves perception in degraded illumination by combining visible texture with thermal contrast. Small objects remain difficult because their discriminative responses are weak, spatially sparse, and sensitive to cross-modal displacement. We present FGCFF-Net, a lightweight frequency-guided fusion detector for visible–infrared small object detection. The framework casts fusion as progressive evidence selection. A lightweight channel pruning module suppresses redundant dual-stream responses; frequency-guided alignment estimates residual spectral compensation; a cross-modal alignment and small-target framework predicts bidirectional spectral weights and deformable offsets; and a frequency-aware cross-modal aggregator conditions channel-spatial attention on frequency discrepancy. The training objective further couples dense detection loss with spectral consistency, offset smoothness, and small-target emphasis. On FLIR, M3FD, LLVIP, and KAIST, FGCFF-Netachieved 92.4% mAP@0.50 on FLIR, 90.7% mAP@0.50 on M3FD, 98.5% AP@0.50 on LLVIP, and 5.72% log-average miss rate on KAIST, with 14.8M parameters and 68 FPS. Together, the results show that frequency-aware alignment can improve multispectral small-object detection without requiring a large model.

Multimodal large language models and diffusion transformers have made text-to-video generation accessible to film production, advertising and social-media editing, but the assessment of generated video aesthetics remains poorly aligned with how viewers diagnose cinematic failure. A video may be sharp at the frame level while failing because of jitter, implausible object motion or mismatch between the prompt and the evolving event. We introduce CineScope-Fuse, a cross-scale semantic fusion network for cinematic aesthetic assessment of multimodal-LLM-generated videos. The method decomposes video quality into spatial fidelity, temporal stability and prompt–event alignment and then recomposes them through a semantic-aware module, a unified cross-attention network and a cross-scale spatio-temporal fusion block. The design turns the three common failure families of generated videos into explicit learning targets rather than treating them as unstructured regression noise. On the proposed Cine-AIGV benchmark and five public auxiliary benchmarks, CineScope-Fuse achieved the strongest overall correlation with human ratings, reaching 0.889 SRCC and 0.884 PLCC on Cine-AIGV. Ablations further showed that semantic alignment learning, cross-dimensional attention and multi-scale temporal fusion contributed complementary gains. These results indicate that cinematic video assessment benefits from jointly modeling low-level fidelity, temporal continuity and high-level narrative alignment, especially for synthetic clips whose visual realism and semantic plausibility fail in different ways.

Visible–infrared detection improves perception in degraded illumination by combining visible texture with thermal contrast. Small objects remain difficult because their discriminative responses are weak, spatially sparse, and sensitive to cross-modal displacement. We present FGCFF-Net, a lightweight frequency-guided fusion detector for visible–infrared small object detection. The framework casts fusion as progressive evidence selection. A lightweight channel pruning module suppresses redundant dual-stream responses; frequency-guided alignment estimates residual spectral compensation; a cross-modal alignment and small-target framework predicts bidirectional spectral weights and deformable offsets; and a frequency-aware cross-modal aggregator conditions channel-spatial attention on frequency discrepancy. The training objective further couples dense detection loss with spectral consistency, offset smoothness, and small-target emphasis. On FLIR, M3FD, LLVIP, and KAIST, FGCFF-Netachieved 92.4% mAP50 on FLIR, 90.7% mAP50 on M3FD, 98.5% AP50 on LLVIP, and 5.72% log-average miss rate on KAIST, with 14.8M parameters and 68 FPS. These results indicate that frequency-aware alignment provides an efficient route for robust multispectral small-object perception.

Robust object detection for autonomous driving requires perception models that remain reliable when visible imagery is degraded by darkness, glare, rain, fog, motion blur, or long-range small targets. Visible and thermal infrared cameras provide complementary evidence, yet many RGB–thermal detectors fuse modalities, mainly as aligned tensors, and may underuse relational structure in channel responses, spatial layouts, semantic scales, and modality-specific uncertainty. This paper presents TopoGraph-Fusion, a hierarchical graph-guided dual-modal object detector that formulates fusion as topology-aware reasoning rather than direct feature concatenation. The proposed framework builds a dual-stream backbone for RGB and thermal images, constructs channel-wise topology through a channel-topology graph aggregation module, derives relation-aware spatial and channel global attention from affinity graphs, and replaces fixed feature-pyramid communication with a Graph-Guided Feature-Pyramid Network. A topology-regularized detection objective further encourages stable cross-modal correspondence while suppressing noisy all-to-all connections. Experiments on M3FD, FLIR, RGBTDronePerson, and VEDAI512 cover road scenes, adverse illumination, drone–person perception, and aerial vehicle detection. Within this validation scope, the results and visual analyses indicate that topology-guided fusion improves small-object recall, cross-modal consistency, and robustness under modality imbalance.

Low-altitude unmanned-aerial-vehicle imagery exposes object detectors to a distinctive combination of tiny object footprints, dense instance layouts, abrupt scale variation, and weak texture under complex urban backgrounds. Existing detectors usually address these factors by adding larger backbones, denser feature pyramids, or heavier attention, but such independent additions often amplify background responses and dilute the fine localization cues needed by small targets. This paper proposes DFA-Det, a dynamic feature augmentation detector that treats low-altitude small-object detection as a coupled problem of context preservation, scale calibration, and content-aware refinement. The method first introduces a poly-kernel inception enhancement branch to preserve shallow structural details while expanding the effective receptive field. It then builds a Multi-Scale Interaction Encoder with Adaptive Feature Prior Learning and an Adaptive Feature Scaling Layer, where the latter contains a Bi-directional Channel Fusion Module that learns channel-wise evidence exchange between adjacent resolutions. Finally, a Hierarchical Refinement and Adaptive Fusion Module performs dynamic upsampling, semantic refinement, and adaptive fusion before the detection decoder. Experiments on the public VisDrone and CODrone benchmarks show that DFA-Det improves small-object precision, crowded-scene recall, and cross-scale robustness compared with representative two-stage, one-stage, transformer-based, and recent YOLO-family detectors. Extensive ablations, heatmaps, and qualitative comparisons indicate that the proposed modules cooperate as a coherent dynamic feature enhancement mechanism rather than isolated architectural attachments.

Ship detection in remote sensing images is crucial for maritime surveillance, port management, and national security. However, existing detectors struggle with complex harbor backgrounds, large-scale variations, and fine-grained inter-class similarities among diverse ship categories. In this paper, we propose SPFDet, a CLIP text-prior-guided structure-enhanced detector that systematically addresses these challenges by integrating vision-language semantic priors with channel-selective feature enhancement. First, a Semantic Prior Component (SPC) employs a frozen CLIP text encoder to generate category-level semantic embeddings from textual descriptions of ship types, which are combined with visual scene priors and detail-aware priors to provide hierarchical domain-specific guidance. Second, a Structure-Enhanced Transformer Module (SETM), equipped with a Cross-level Channel Selection Module (CCSM) and Multi-Head Cross-Attention (MHCA), selectively enhances discriminative channel responses associated with hull contours, deck textures, and fine structural patterns across encoder layers. Third, a Fused Category-Dominated Feature (FCDF) generation mechanism integrates CLIP-based semantic priors with multi-scale visual features through category-guided attention, producing category-aware representations for precise classification and localization. We further introduce a Category-Semantic Consistency Loss to enforce alignment between predicted features and CLIP text priors. Extensive experiments on HRSC2016 and ShipRSImageNet benchmarks demonstrate that SPFDet achieves 97.2% and 72.8% mAP respectively, providing consistent improvements over strong detection baselines while maintaining competitive inference speed.