Peiguang Jing

In semantic recognition based on surface electromyography (sEMG), low signal-to-noise ratios and physiological differences across subjects limit the ability of unimodal sEMG features to fully represent speech semantics, constraining recognition accuracy. Existing multimodal fusion methods can improve accuracy but often require large models and high computational complexity. Mismatches in feature spaces and learning capacity between multimodal teachers and unimodal students further limit cross-modal knowledge transfer. To address these issues, this paper proposed cross-modal adaptive feedback knowledge distillation (CAFKD). The multimodal teacher model uses Bi-GRU to fuse temporal features from synchronously acquired audio, lip image sequences, and sEMG signals, generating high-level representations that encode complementary semantic information. This multimodal knowledge was transferred to an sEMG-only student through a joint loss combining classification, output distribution distillation, and intermediate feature distillation. In addition, the cross-modal adaptive feedback mechanism (CAF) periodically adjusted the teacher using stagewise output feedback from a temporary student, enabling teacher guidance to better match the student’s learning state. In subject-independent evaluation on the AVE-Speech dataset, CAFKD achieved 78.13% accuracy on the 101-class closed-set short-utterance classification task, exceeding the unimodal Bi-GRU baseline by 4.82 percentage points. Ablation experiments showed that adding intermediate feature distillation to output distribution distillation increased test accuracy by 1.49 percentage points. CAF further improved accuracy by 1.46 percentage points over joint distillation. The student model also retains the same parameter count and FLOPs as the Bi-GRU baseline, at 4.42 M and 0.29 G, respectively. These results indicate that CAFKD improves recognition performance on unseen subjects while preserving a lightweight architecture for sEMG-only inference.

Referring expression segmentation (RES) aims to segment the visual region described by a natural-language expression, providing a pixel-level language-vision interface. However, mainstream RES architectures keep the visual encoder query-agnostic and defer cross-modal interaction to a downstream stage, so target-relevant evidence is diluted by background activations before guidance can take effect, while alignment, re-weighting, and refinement pile into one serial stage that inflates inference latency. We propose the Subject-Aware Visual Network (SAVN), a zero-stage RES architecture that performs cross-modal interaction inside the visual encoder and produces masks without any decoder. A frozen vision-language text encoder supplies layer-wise end-of-text (EOS) features as query semantics, and Cross-Modal Feature Focusing Modules (CMFMs) inserted at multiple depths of a pre-trained Vision Transformer use them to modulate visual features through two complementary paths: spatial re-weighting of tokens by query relevance, and semantic injection that reinforces query-described attributes. The attention maps of the last CMFM are converted into the segmentation mask by a lightweight convolutional read-out layer, and a terminal-constrained strategy applies pixel-level supervision to this final CMFM only, so the earlier CMFMs learn focusing behavior implicitly. Experiments on RefCOCO, RefCOCO+, and RefCOCOg demonstrate that this decoder-free design matches or exceeds every compared method on all eight evaluation splits while running faster than representative decoder-based methods and preserving accuracy under reduced-precision inference, suggesting query-adaptive encoding as a viable alternative to decoder-based pipelines. A controlled ablation under an identical backbone and protocol attributes an 8.2 mIoU gain to the focusing mechanism rather than to the pre-training prior.

Underwater visual degradation makes it difficult for the image modality to represent video semantics, while label sparsity in underwater scenes leads to weak inter-category correlations, thereby degrading the performance of multi-label classification. To address these issues, this paper proposes a Semantic-Enhanced Underwater Videos Multi-Label Classification Network Based on Structural Graph Convolution (SEMGCN). Specifically, the proposed method first disentangles the image and text modalities into shared and private representations, and enhances feature representation capability through orthogonal constraints and feature reconstruction. Moreover, a Cross-Modal Category-Aware Module (CCAM) is constructed to model interactions between image and text features and perform bidirectional cross-attention with category-label text embeddings, thereby generating category-aware initial node representations. Furthermore, to alleviate the limitation of semantic propagation caused by sparse label co-occurrence, a Structural Graph Convolutional Network (SGCN) is proposed. By integrating explicit co-occurrence relationships with implicit structural similarity relationships, the proposed model collaboratively captures both explicit and latent semantic associations, thereby improving multi-label classification performance under label-sparse conditions. Experiments were conducted on the self-constructed Underwater Video Multi-label Classification Dataset (UVMC) and the public MLSV2018 dataset. The experimental results show that SEMGCN achieves Average Precision scores of 0.8645 and 0.8388 on UVMC and MLSV2018, respectively, demonstrating the effectiveness of the proposed method.

The explosive growth of micro-videos has intensified the demand for efficient hash-based retrieval techniques, where the inherent co-occurrence of semantic concepts necessitates effective modeling of multi-label correlations. However, label confusion caused by high-density information, along with the complex entanglement of multi-label dependencies, still poses significant challenges to generating robust hash codes. To address these issues, we propose a proxy-guided contrastive hashing network (DCPH), which integrates contrastive learning with proxy guidance to generate semantic-constrained hash representations for micro-videos. Specifically, to handle the frequent yet tricky visual differences, we design a heterogeneity-aware visual-to-semantic projection (HAVSP) component. HAVSP aggregates overall visual representations in the shared space via cross-modal cyclic projection, and introduces dual-constrained contrastive alignment to alleviate semantic ambiguity between visual features and textual prompts. To explicitly model multi-label dependencies, we develop a proxy-guided decompositional hashing (PGDH) component. PGDH dynamically decomposes label semantics into sequential binary codes via serial hash mappings, and incorporates multi-label proxy-center learning to optimize the commonality and mutual exclusivity among entangled micro-videos. Comprehensive experiments on a large-scale benchmark dataset demonstrate that DCPH achieves state-of-the-art retrieval performance, verifying its efficiency and accuracy in diverse micro-video retrieval.

Obtaining large amounts of paired data is difficult in the field of bit-depth enhancement (BDE), which refers to the enhancement of low-bit depth (LBD) into high-bit depth (HBD) data to provide better visualisation. Existing BDE algorithms, whether traditional or learning-based, all use direct quantization to simulate real LBD data. However, the real degradation process is more complicated, leading to significant performance drops when these algorithms are applied to real LBD images. Therefore, we propose a novel real-world degradation model which simulates the degradation process from HBD to LBD, and a lightweight Real-world Multi-Scale BDE model (Real-MSBDE) with great generalization ability when dealing with real LBD data, both of which together accomplish the real-world BDE task. It is worth noting that this degradation model is not a regression model commonly used in degradation processes of other fields, but a classification-based model designed according to the statistical characteristics of real data. This model enables achieve the purpose of constructing paired LBD-HBD data. Using the paired data generated by proposed degradation model for training, Real-MSBDE model can more accurately capture the transformation from real LBD images to real HBD images than those BDE algorithms trained with quantized LBD images. Experiments demonstrate the effectiveness of proposed classification-based degradation model and Real-MSBDE model. Experiments also show that the data generated by classification-based degradation model can assist other learning-based BDE algorithms in improving their performance on real LBD data, which further validate the superiority of proposed degradation model.

The Linfen–Yuncheng Basin is located on the southern edge of the Fenwei Fault Zone, influenced by intense tectonic activity, thick Quaternary sedimentation, and anthropogenic disturbance, it exhibits prominent characteristics of ground subsidence and fissure development. However, uncertainties still exist regarding the primary controlling factors of subsidence. This study employs multi-temporal InSAR data, combined with small baseline subset (SBAS–InSAR) technology to invert the high-precision ground line of sight deformation fields, and conducts time-series decomposition analysis using the Seasonal Trend Decomposition (STL) method. The results show that from 2017 to 2025, subsidence was mainly concentrated in the central and southern regions of the basin, with a maximum cumulative subsidence exceeding 200 mm and an average annual subsidence rate of −40 mm/year. Its spatial distribution is highly consistent with major structural zones such as the Zhongtiao Mountain Front Fault and the Linyi Fault, indicating that fault activity exerts a significant controlling effect on subsidence patterns. Groundwater level fluctuations are positively correlated with overall ground subsidence, and the response rate of different monitoring points is constrained by differences in aquifer depth and permeability. Groundwater aquifer points exhibit rapid and reversible subsidence response, while confined aquifer points are affected by low-permeability or compressible layers, showing a significant lag effect. The research results indicate that time-series analysis based on InSAR can not only effectively reveal the subsidence evolution process at different scales, but also provide a scientific basis for groundwater resource regulation, geological disaster prevention and control, and sustainable regional land utilization.