
Federated incremental semantic segmentation aims to design privacy-preserving collaborative training methods to continuously segment new classes from distributed datasets. However, extending this task to the abdominal and pelvic domain introduces new challenges arising from the inherent characteristics of abdominal and pelvic CT images, including low foreground-background contrast and pixel-level data imbalance. To address these issues, we propose a novel MedFISS, an abdominal and pelvic CT image-oriented federated incremental semantic segmentation framework. Specifically, we leverage textual prior knowledge as shared cross-client and cross-task semantic guidance. Based on this prior, we design a task-relevance guided adaptive pseudo labeling mechanism and a structural consistency constraint, which improve the reliability of pseudo labels in low-contrast abdominal and pelvic CT images and alleviate catastrophic forgetting. Additionally, we propose mitigation strategies at the local training stages to progressively reduce the bias introduced by pixel-level imbalance in federated incremental semantic segmentation. Furthermore, we quantify the information dilution of different classes during model aggregation and apply a weighting mechanism to mitigate the varying degrees of information loss across classes. Finally, we conduct experiments on four public datasets to validate the effectiveness of our method. Our method achieves a 5.88% improvement in the final mIoU compared with the current model on a cross-task dataset, i.e., AMOS-LITS. The code is available at: https://github.com/hmt5912/hmt.
Automated surface-crack inspection provides observable evidence for facility maintenance, but visible crack patterns cannot directly reveal the internal 3D fissure morphology that controls materials degradation and permeation-driven durability loss. This paper proposes Crack2Fissure, a physics-guided conditional Diffusion framework that formulates depth-wise fissure evolution as a spatiotemporal slice sequence, embeds inter-slice transition physics through optical-flow information, and generates structurally coherent 3D fissures under supervision from real fissure data. The framework is trained with paired surface-crack and X-ray micro-computed tomography data from Engineered Cementitious Composite specimens. The generated fissures reproduce geometry- and permeability-related indicators, with relative errors of 0.59% for tortuosity, 4.29% for roughness, 3.16% for fissure width, and 3.70% for flow rate. Crack2Fissure therefore extends automated surface inspection toward probabilistic subsurface damage assessment while retaining explicit applicability.
Backdoor attacks play a crucial role in assessing model robustness and privacy risks. However, existing multimodal backdoor methods often fail in high-level downstream tasks such as visual emotion analysis (VEA), as they overlook the issue of semantic asynchrony between different modalities, resulting in suboptimal emotional perception. To address this limitation, we propose EmoAttack, a unified emotion-oriented backdoor attack framework that dynamically optimizes tunable prompts and fine-grained triggers through dual latent space modeling. By exploiting the correlation between emotional semantics and the attack mechanism, our approach optimizes these prompts across different training phases to inject fine-grained target-aware knowledge. As a result, EmoAttack effectively decouples prompt injection from task-specific transfer, facilitating precise emotional cue mining while preserving semantic consistency across diverse visual contents. Extensive experiments on seven public VEA datasets ( 166K images) show that EmoAttack achieves both high stealthiness and stability, with an average clean accuracy improvement of 11.6% and attack success rate (ASR) up to 89.3% even under the CleanClip defense, outperforming state-of-the-art methods.
The rotating triboelectric nanogenerator (R-TENG) has emerged as a core technology in micro-energy harvesting and micro-signal detection. However, the high impedance, high-voltage and low-current output characteristics of R-TENG structures, lead to bottlenecks such as impedance mismatch, difficult topological adaptation and weak anti-interference capabilities. Under controlled bench-test conditions with 1-3 Hz mechanical excitation, customized rectification and energy storage circuits improve R-TENG energy harvesting efficiency from 60% to 70%-95%, a 10%-35% gain primarily by mitigating impedance mismatch. Sensing accuracy for rotational and wind speed is +/- 0.5%-+/- 2% of full scale, varying with methods and environmental interference. This paper systematically reviews the research progresses of the micro-energy conversion and micro-signal processing technologies for R-TENG. Regarding micro-energy harvesting, the key technologies such as rectification, energy storage optimization, and DC-DC buck conversion are analyzed systematically. For micro-signal processing, indepth analysis is conducted on signal processing methods such as filtering, noise reduction, and feature extraction in typical application scenarios such as motor speed estimation and wind speed measurement. The core contributions include: established a comprehensive analysis framework that bridges theoretical foundations and engineering applications; highlighted the performance metrics and practicality through structured comparisons; identified the research gaps in circuit stability, model generalizability and adaptation to extreme environments. Finally, this article analyzes the development directions of R-TENG such as intelligent adaptability, anti-interference ability, and universal modularization, which will help promote the innovative application of RTENG in micro-energy harvesting and micro-signal processing.
A nonconforming virtual element scheme for the Allen-Cahn equation on polygonal meshes is developed. The numerical scheme employs a variable-time-step two-step backward differentiation formula (BDF2) for temporal discretization and a lowest-order nonconforming virtual element method in space. With the help of discrete we obtain the error estimates of temporal semi-discrete system under a mild time-step ratio 0 < r(k) < r(max) approximate to 4.8645. By using temporal-spatial error splitting approach, we establish the L-infinity-norm bound of discrete solution, and the unconditional optimal L-2-error estimates are carried out without CFL condition. Numerical experiments validate the theoretical findings across diverse polygonal meshes.