Xiaoyang Yuan

Fragmented object segmentation in optical remote sensing imagery is a challenging visual information processing problem because target regions often exhibit large scale variation, discontinuous spatial layouts, and strong background ambiguity. Landslide extraction is a representative case of this problem, where collapsed areas may appear as scattered but semantically related patches and can be easily confused with barren soil, rocks, roads, and shadows. To address these challenges, this paper proposes a Target-Aware Adaptive Multi-scale Information Reconstruction Network, termed GTAR-Net. The proposed framework converts target-specific visual characteristics into learnable information modeling mechanisms. First, an Orthogonal Gated Context Bottleneck (OGCB) models directionalcontextual information at the deepest encoder layer to enhance structural continuity for fragmented regions. Second, a Dynamic Multi-scale Adaptive Module (DMAM) performs scene-aware information routing across hierarchical encoder features, enabling adaptive aggregation of both large target bodies and small collapse scars. Third, a Target-Guided Progressive Reconstruction Decoder (TGPRD) uses high-level semantic information to recalibrate skip features, suppress background interference, and recover weak boundaries. Experiments on the CAS Satellite, CAS UAV, and Bijie datasets demonstrate that GTAR-Net achieves superior robustness compared with representative segmentation methods. The codes and results are available at https://github.com/AmazingJ-123/landslideseg-GTARNet/.

Hybrid brain–computer interfaces (BCIs) have attracted growing research attention because they combine the millisecond-level temporal resolution of electroencephalography (EEG) with the spatially informative hemodynamic responses of functional near-infrared spectroscopy (fNIRS). However, most existing deep fusion methods rely on static late-fusion strategies, which tend to underexploit latent cross-modal dependencies and are vulnerable to modality-specific signal degradation. To address these limitations, we propose MGFNet, a multi-granularity fusion network for hybrid BCI decoding. MGFNet contains three components: (1) intra-modal encoders that learn modality-specific spatiotemporal representations from EEG, oxygenated hemoglobin (HbO), and deoxygenated hemoglobin (HbR) signals; (2) cross-modal interaction encoders that temporally align paired modalities and use dilated convolutions to capture long-range EEG-fNIRS dependencies; and (3) a Coupling-Guided Sparse Component Routing (CGSCR) module that estimates sample-specific cross-modal coupling and performs adaptive discrete routing. We further introduce a deep supervision strategy to stabilize optimization and improve branch-level discriminability. Under a within-subject held-out evaluation protocol on a public benchmark dataset, MGFNet achieved classification accuracies of 99.40% on the n-back task and 99.03% on the word generation (WG) task, outperforming representative comparison methods evaluated under a matched protocol. Ablation studies further confirmed the contributions of the intra-modal encoders, the cross-modal interaction encoders, and the CGSCR module. Under controlled EEG corruption with additive white Gaussian noise at −10 dB, MGFNet outperformed a static-fusion variant by 9.23 percentage points on the n-back task and 6.31 percentage points on the WG task. These results support the effectiveness of MGFNet in the present offline within-subject setting and indicate improved robustness under controlled single-modality degradation.

Infrared small target detection constitutes (IRSTD) a critical capability in remote surveillance and tracking systems, serving essential roles in applications such as early warning, reconnaissance, and target monitoring. However, existing IRSTD methods primarily rely on attribute-specific feature modulation (i.e., spatial and channel features modulate their own responses) or feature-driven attention in transformer-style designs, computed directly from input features, lacking explicit query-guided mechanisms to coordinate spatial and channel information, which limits robust small target representation. To this end, we propose CSQNet, a query-driven channel-spatial attention network that leverages learnable queries to explicitly guide feature selection and attention generation. At its core, the Residual Channel-Spatial Query (RCSQ) Attention mechanism constructs decoupled yet complementary pathways along channel and spatial dimensions, enabling cross-dimensional interaction for enhanced target discrimination. Based on RCSQ, a query-guided encoder-decoder framework is designed, in which the Progressive Query-Guided Encoder progressively refines feature representations, and the Residual Query-Aware Decoder aggregates multi-scale query-enhanced features for precise target reconstruction. Experiments on three public datasets demonstrate that CSQNet achieves superior segmentation accuracy and robustness in cluttered infrared scenes.

RGB-T salient object detection (SOD) integrates visible and thermal cues for robust localization under complex conditions. However, existing methods often employ indiscriminate fusion strategies that assume equal reliability across modalities. This approach fails to dynamically mitigate interference when one modality degrades (e.g., thermal crossover), leading to the propagation of noise and the corruption of reliable features. To address this, we propose the Conditional Multi-scale Cross-Modal Adaptive Fusion (CMCAF) network. The core philosophy of CMCAF is to treat the thermal modality not merely as a feature source, but as a dynamic condition to modulate RGB processing. Specifically, a Shared Swin Backbone (SSB) is employed to extract aligned cross-modal representations. At the bottleneck, a Thermal-Conditioned Modulation (TCM) block generates channel-wise affine parameters. It functions as a gate: amplifying reliable cues when thermal data is salient, while effectively suppressing noise propagation to protect RGB semantics when thermal data is unreliable. To accommodate object scale variations, a Scale-Aware Fusion (SAF) module acts as a scale arbitrator, adaptively balancing semantic context and fine details. Furthermore, a Thermal-Guided Gating Decoder (TGGD) screens skip connections via dual gating, filtering out background noise backflow during reconstruction. Extensive experiments on RGB-T and RGB-D benchmarks demonstrate that CMCAF consistently outperforms state-of-the-art methods, exhibiting superior accuracy and strong robustness against modal noise. The codes and results can be accessed at https://github.com/AmazingJ-123/RGBT-SODCMCAF