Search Results (129)

High-resolution remote sensing object detection is a fundamental task in Earth observation. However, the storage, transmission, and downstream processing of high-resolution images impose substantial bandwidth, memory, and computational burdens. Compressed sensing (CS) samples and compresses the signals simultaneously, thereby reducing data transmission and storage overhead. Unfortunately, most existing CS-based pipelines require explicit image reconstruction before downstream inference, leading to heavy computational overheads and poor scalability for high-resolution remote sensing images (RSIs). This work focuses on post-acquisition image compression and explores algorithmically, rather than from a physical hardware implementation perspective, whether explicit image reconstruction is an indispensable intermediate step prior to object detection. To this end, we propose HR²SIOD-CL, an end-to-end compressed learning (CL) framework that performs object detection directly on CS measurements of high-resolution RSIs without explicit image reconstruction. HR²SIOD-CL integrates an entropy-driven content-aware adaptive sampling strategy and a measurement-domain detection backbone for multi-scale feature extraction. Although jointly optimized during training, the adaptive sampling and detection modules can be decoupled for flexible deployment. For fair comparison, an extra lightweight reconstruction network equipped with a single-step data-consistency correction is constructed as the baseline. Extensive experiments on the NWPU VHR-10 and DIOR datasets show that across various sampling ratios, HR²SIOD-CL surpasses the reconstruction-based detection method when integrated into two-stage detectors, and achieves comparable or superior detection performance to the reconstruction-based counterparts when integrated into single-stage detectors. Meanwhile, its computational overhead and GPU memory consumption are merely 6.97% and 35.74% of those of the reconstruction-based counterpart, respectively. These results indicate that CS measurements can function as an effective intermediate representation for object detection, and explicit image reconstruction is not a prerequisite when detection is the primary objective. Full article

The quality and safety supervision of agricultural products at the grassroots level has long faced the triple superposition dilemma of small-sample sampling, multi-source evidence conflict, and risk chain evolution. Although existing data-driven models have considerable accuracy, they are difficult to leverage for intervention decisions, and the simple serial connection of traditional Bayesian networks and evidence theory cannot respond to dynamic scenarios. Aiming at this research gap, this paper constructs a dynamic risk assessment model, CIBE-DR, that deeply couples Bayesian networks with evidential reasoning. It contains three core innovations. First, the structure learning method of the causally identifiable Bayesian network embeds a graded do-calculus identifiability score covering both back-door and front-door criteria into the BDeu scoring function and combines this reward with an expert-prior divergence penalty that breaks Markov equivalence so as to realize the transition from relevance modeling to intervention decision modeling. Second, the conflict-aware adaptive evidence synthesis rule orthogonally decomposes multi-source conflict into an epistemic component and an ontological component, which are modeled respectively by Tsallis belief entropy and abductive inference over a discrete twenty-seven-point heterogeneity hypothesis space and are then fused under a reparameterized Dempster–Yager interpolation in which the two endpoints recover the two named rules under a single consistent interpretation. Third, the bidirectional closed-loop coupling mechanism between BN and ER realizes the mutual calibration between the conditional probability table and the evidence credibility prior under a Lyapunov monotone descent argument with the explicit Lipschitz bound $L_{\theta }$ ≤ 0.028 < 1, endowing the model with time-varying self-correction ability. Based on experiments on 156,847 sampling samples from counties and townships in East China, Central China, and Southwest China from 2021 to 2024, the proposed method achieved the best value in six of the seven evaluation indicators, with a minority recall of 0.864 ± 0.014, an intervention effect estimation error of 0.063 ± 0.005, and a dynamic response delay of 2.8 ± 0.3 days, significantly ahead of eleven mainstream baselines under the McNemar test on classification ($p$ < 0.001) and the Wilcoxon signed-rank test on intervention-effect estimation ($p$ < 0.001). The only indicator on which CIBE-DR does not lead is overall accuracy, which is 0.002 lower than that of Transformer; this difference does not reach statistical significance under the McNemar test ($p$ = 0.32) and does not weaken the value of grassroots supervision in the strong-imbalance scenario where the positive rate is only 1.04%. The robustness advantage of the model is particularly prominent in the scenarios of sparse data, adversarial perturbation, and prior-graph incompleteness, and the intervention-effect estimates were additionally validated against two post-2022 policy interventions with absolute deviations of 1.4 and 1.2 percentage points respectively. These results verify the product gain and grassroots deployability of the three mechanisms. Full article

High-precision urban vegetation mapping is essential for assessing carbon sink capacities, mitigating the urban heat island effect, and supporting sustainable development. Although deep learning and high-resolution remote sensing have advanced automated vegetation monitoring, existing models still face challenges when a common segmentation architecture is evaluated under different imaging geometries. In this study, Cityscapes and ISPRS Vaihingen are treated as two independent benchmarks representing perspective street-level imagery and orthographic aerial imagery, rather than as simultaneous cross-view inputs. “Background dominance” caused by perspective distortion and the “gridding artifacts” inherent in orthographic textures severely constrain segmentation accuracy across varying vegetation scales, particularly for small targets. To address these limitations, we propose a Scale-Aware Mixture of Experts (SA-MoE) architecture for fine-grained vegetation segmentation under two distinct imaging views, together with a scene-specific optimization strategy. The core SA-MoE framework consists of two main components. First, the spatial gating network uses a temperature polarization mechanism with τ = 0.5 to adjust the initial logit maps, sharpening expert-weight differences while preserving stable gradient propagation. Second, we use a heterogeneous expert group with five parallel branches: a pixel-level expert, three spatial experts with different dilation rates, and a global average-pooling expert. A dynamic pixel-level weighted fusion mechanism is then applied, decoupling feature extraction from receptive-field allocation. Furthermore, to address the heterogeneity of “hard samples” and “label noise” across the two benchmark settings, we introduce a scene-specific optimization strategy. Our findings show that the Focal-Dice (FD) loss is more suitable for perspective scenes with severe target imbalance and hard-to-classify vegetation targets, whereas the Cross-Entropy (CE) loss is more robust to boundary jitter in orthographic imagery. Comparative experiments on the Cityscapes (perspective view) and ISPRS Vaihingen (orthographic view) datasets reveal that SA-MoE achieves a highly competitive balance between computational efficiency and fine-grained segmentation, particularly in micro-target recall. Notably, the recall for extra-small (XS) scale targets in the aerial dataset improved by 3.21 percentage points compared to the second-best model. For the street-level dataset, our model achieved competitive global performance in terms of Overall Accuracy (OA), Precision, and F1-Score. However, we also observed a performance trade-off, where Transformer-based models maintained an advantage in preserving fine boundary details for these extra-small targets. In the routing analysis, we observed a pattern that we refer to as “receptive field inversion”, in which the model assigns lower weights to large-dilation experts for large canopy regions in orthophotos. We interpret this pattern as a plausible routing hypothesis. Overall, SA-MoE offers an efficient and adaptive solution for urban vegetation mapping under two imaging views. Full article

Indoor positioning using commodity Wi-Fi infrastructure is attractive for smart buildings and Internet of Things applications, but the practical accuracy of Wi-Fi Fine Timing Measurement (FTM) remains limited by non-line-of-sight propagation, multipath delay, access-point-dependent ranging bias, and unstable anchor geometry. This paper proposes a measurement-reliability learning and geometry-constrained fusion framework, termed MRL-GCF, for robust horizontal Wi-Fi FTM indoor localization. MRL-GCF learns the reliability of each access-point observation from a multi-factor representation that includes Received Signal Strength Indicator (RSSI), logarithmic FTM range, short-window range stability, RSSI fluctuation, access-point visibility, abnormal-range tendency, and coarse anchor geometry. A lightweight heteroscedastic neural calibrator estimates both range bias and observation uncertainty. A supervised reliability-regime head is further trained from residual-regime soft targets, and its entropy is used as a propagation-ambiguity measure. The learned uncertainty is fused with propagation ambiguity, map obstruction, material-aware obstruction cues, and anchor geometry to select reliable anchors and construct a trust-weighted nonlinear least-squares localization objective. To avoid overestimating performance from repeated scans at identical survey points, both scan-level and point-held-out protocols were adopted. Experiments were conducted in a lobby, a classroom, and a dormitory using 4410 synchronized RSSI-FTM scans. On 882 scan-level test queries, MRL-GCF achieved mean absolute errors of 0.88 m, 0.55 m, and 1.20 m, with sub-3 m success rates of 98.0%, 99.0%, and 96.5%, respectively. Additional replay-based dynamic, temporal, cross-device, AP-density, uncertainty-calibration, map-availability, and coefficient-sensitivity analyses were included to examine deployment-oriented robustness. These results indicate that learning measurement reliability while preserving geometric constraints provides a practical and interpretable solution for robust Wi-Fi FTM indoor positioning. Full article

Three-dimensional (3D) point cloud change detection is essential for urban monitoring and environmental analysis, yet existing methods mainly rely on point-wise semantic differences and overlook change–unchange boundaries and object edges—key geometric cues for precise localization, especially for subtle or gradual changes. To address this, we propose GSANet, a Geometric Structure-Aware Siamese Network that explicitly integrates boundary and edge priors into sampling and feature learning. First, a Boundary-Aware Subsampling (BAS) strategy preserves key points near change boundaries while reducing redundancy, and a Boundary-Aware Binary Cross-Entropy (BA-BCE) loss assigns higher supervision weights to boundary points, enhancing learning in ambiguous regions. Second, an Edge-Aware Siamese Network captures robust local shapes by embedding edge priors into feature extraction, incorporating Edge-Aware Adaptive Graph Convolution, Edge-Aware Downsampling, and Cross-Attention Upsampling to maintain structural consistency across temporal branches. Additionally, a Difference Enhancement Module (DEM) amplifies feature discrepancies between bitemporal point clouds, improving sensitivity to subtle changes. Extensive experiments on a street-level dataset and urban dataset show our method outperforms state-of-the-art approaches. Full article

The rapid growth of Internet of Things (IoT) devices generates high-dimensional, high-velocity data streams that demand real-time machine learning (ML) inference under strict hardware constraints. We propose the Adaptivecoreset Selection Engine (ACS-Engine), a unified framework for adaptive, differentiable, and resource-aware coreset selection on streaming IoT data. ACS-Engine introduces three tightly integrated innovations: (i) Differentiable Greedy Sampling (DGS), which relaxes discrete subset selection via Gumbel-Softmax reparameterization to enable end-to-end gradient-based optimization; (ii) Entropy-Aware Regularization (EAR), which promotes coreset diversity and provides implicit concept drift detection through a self-calibrating entropy threshold; and (iii) Resource-Aware Memory Management (RAMM), which dynamically adjusts the target coreset size based on real-time hardware telemetry—available memory, CPU utilization, remaining energy, and sampling frequency. Evaluated on eight real-world IoT datasets spanning three heterogeneous edge platforms, ACS-Engine achieves 15× memory reduction and a 20% energy efficiency improvement while retaining 98% of full-dataset accuracy, with a per-sample latency of 2 ms that satisfies real-time edge deployment requirements. Full article

Deep neural networks (DNNs) have shown outstanding performance in visual recognition tasks within vision sensor networks; however, they are still vulnerable to adversarial manipulations and imperceptible perturbations that can lead to erroneous predictions. To address that, this paper presents AEGIS, a semantic aware and uncertainty guided adversarial detection framework designed for robust image classification in vision sensors pipelines. At its core, a SemantiGAN module functions as a multi-class semantic discriminator, identifying and filtering visually inconsistent adversarial inputs before they propagate further in the pipeline. For inputs that pass this stage, a stochastic augmentation process generates test time variations, from which handcrafted instability metrics FlipScore, Prediction Inconsistency, Layerwise Cosine Similarity (early and mid layers), and Entropy are computed. These features are aggregated into a compact five dimensional vector and processed by an Evidential Deep Learning (EDL) classifier, which models output evidence using a Dirichlet distribution to yield both class predictions and calibrated uncertainty estimates. Evaluations on the Tiny ImageNet dataset across six categories clean, FGSM, PGD, patch-based, functional, and geometric attacks demonstrate the effectiveness of AEGIS. The proposed framework achieves an AUROC of 92.1%, an AUPRC of 90.2%, and an accuracy of 90.7%, outperforming conventional softmax-based detectors in terms of detection performance, robustness, interpretability, and uncertainty calibration. Full article

Redundant channels, sparse electrode topology, and insufficient cross-layer feature fusion limit electroencephalography (EEG)-based emotion recognition. This study proposes the Topology-Aware Dual-Bridge Fusion Network (TDBF-Net), a compact framework that integrates adaptive channel selection, topology-aware sparse convolution, and bidirectional bridge fusion. First, sample entropy, dispersion entropy, and fuzzy entropy are fused to estimate channel importance, while particle swarm optimization (PSO) learns the entropy weights and an elbow-based criterion determines the retained channel subset. Second, differential entropy (DE) features from the $\theta$, $\alpha$, $\beta$, and $\gamma$ bands are mapped to an $8\times 9$ sparse topological tensor according to electrode locations. A fixed spatial validity mask is applied before and after convolution to suppress invalid responses from zero-padded regions and preserve real electrode topology. Third, a dual-bridge fusion module recalibrates shallow and deep features in both directions through channel attention and gated fusion, and a bidirectional long short-term memory network (BiLSTM) further captures short-term temporal dependencies. Subject-dependent experiments on the SJTU Emotion EEG Dataset (SEED) and the Database for Emotion Analysis using Physiological Signals (DEAP) show that TDBF-Net achieves $97.62\%$ ± $1.59\%$ accuracy on SEED and $98.46\%$ ± $0.94\%$ and $98.14\%$ ± $0.77\%$ on DEAP valence and arousal, respectively. Paired DEAP ablations support topology and bridge contributions for valence, whereas the corresponding arousal differences are not significant. Selector controls, robustness tests, computational profiling, and held-out visualizations further characterize the method’s compression, cost, and interpretability. The evidence supports TDBF-Net as an effective subject-dependent framework while leaving subject-independent and cross-dataset generalization for future validation. Full article

Oil-well anomaly monitoring supports safe and efficient oil-and-gas production, but delayed recognition of abnormal operating states can reduce lifting efficiency, trigger costly interventions, and increase operational risk. Existing data-driven detectors are also vulnerable to optimistic estimates when segmentation, normalization, and train–test splitting leak information across source files. This study presents ALSTMResNet, an application-oriented active-learning workflow built on an LSTMResNet backbone for one-step-ahead anomaly screening under partial labels. The workflow converts real 3W well records into leakage-aware file-wise splits, train-only standardized sliding windows, and a fixed 15-channel representation that combines process measurements with temporal covariates; active learning is used only during training to select additional labels while leaving the deployed backbone unchanged. Experiments on the retained benchmark split obtain an $F_1$-score of 0.9354, and five repeated file-wise trials give an average $F_1$-score of $0.8661\pm 0.0666$ while reducing retraining time relative to the full-data backbone. Label-budget and acquisition-policy analyses show that the workflow remains competitive under constrained labels, although uncertainty, entropy, margin, least-confidence, and random querying have limited separation in the present binary setting. These results indicate that ALSTMResNet can support cost-aware oil-well anomaly monitoring when labels are partially available and retraining resources are constrained. Full article

Entropy-based soft-min operators and Gibbs policies are widely used in reinforcement learning to smooth greedy decisions and regulate stochastic action selection. Their behavior depends critically on an inverse temperature parameter, which is often chosen through fixed values, annealing schedules, entropy targets, or worst-case bounds. These rules do not directly control the local error induced by soft action selection and do not account for the geometry of the action-value vector. This paper develops a geometry-aware conformal calibration framework for selecting inverse temperatures in discounted finite-action reinforcement learning. The analysis distinguishes the operator-level soft-min approximation error from decision-level Gibbs excess and shows how both quantities depend on local value gaps and a near-optimal action structure. A conformal order statistic rule is then used to obtain finite-sample marginal control of the selected local score under exchangeability. A geometry-conditional extension assigns different temperatures to different action-value geometries. The Bellman analysis clarifies that fixed temperature maps preserve contraction, while data-dependent conformal selectors are best interpreted as post-training action selection rules. A finite-MDP experiment empirically confirms the conformal coverage behavior and shows that calibration must be applied to a decision-relevant cost representation when value estimates contain optimistic value estimation traps. An Atari SpaceInvaders experiment with a frozen deep Q-network shows that the learned action-value geometry is heterogeneous, that decision-level calibration is less conservative than operator-level calibration, and that gated geometry-conditional selection preserves most of the greedy return while providing explicit marginal score control of randomized decision excess with low online overhead. Full article

Cross-condition fault diagnosis is important for ensuring the reliable operation of mechanical systems. However, most existing methods assume that the source and target domains share identical fault label spaces and data distributions, limiting adaptation to unknown faults and cross-condition shifts in practical industrial scenarios. To address this issue, a structured, reliable, domain-specific open-set domain adaptation method is proposed. The proposed method first constructs a domain-specific batch normalization-based feature extraction network, in which independent normalization branches model statistical discrepancies under different operating conditions; it then designs a cross-domain structured representation consolidation module to enhance feature discriminability through source-domain anchor compactness, target-domain multi-view contrastive, and prototype entropy regularization constraints; an open-set boundary learning mechanism is further introduced to establish a discriminative boundary between known and unknown classes; finally, a reliability-aware pseudo-label propagation strategy refines target-domain pseudo-labels and imposes separate prediction-consistency constraints on known and unknown classes. Experimental results on the CWRU bearing dataset and the self-built rolling bearing dataset show that the proposed method achieves average H-scores of 93.32% and 97.65%, respectively, on open-set transfer tasks. Compared with several baseline methods, the proposed method achieves a better balance between known-class recognition and unknown-class detection, thereby improving cross-condition open-set fault diagnosis performance. Full article

Background: Deep-learning systems for endoscopic image classification are commonly evaluated with random data splits, which may overestimate performance under acquisition shift; uncertainty-aware selective prediction may improve reliability by allowing a model to abstain on uncertain cases. Methods: We evaluated binary abnormality detection in two endoscopic imaging domains: laryngeal contact-endoscopy narrow-band imaging (CE-NBI; 210 patients, patient-level) and gastrointestinal endoscopy (HyperKvasir; 6746 images, image-level, predominantly white-light). ImageNet-pretrained ResNet-50 deep ensembles were assessed under a resolution-defined acquisition-shift stress test; for the gastrointestinal data, a random stratified split was additionally used as an in-distribution reference. We evaluated discrimination, calibration, decision-curve analysis, and entropy-based selective prediction. Results: In the gastrointestinal dataset, random-split evaluation produced high performance (AUROC 0.994, 95% CI 0.990–0.996; AUPRC 0.991). Under resolution shift on the same data, performance fell to AUROC 0.723 (95% CI 0.710–0.736; AUPRC 0.690; sensitivity 0.417; specificity 0.874); the two intervals do not overlap. Selective prediction improved reliability among retained cases: under resolution shift, accuracy rose from 0.683 at full coverage to 0.852 (95% CI 0.835–0.871) at 25% coverage (balanced accuracy 0.646 → 0.798). Predictive entropy was significantly higher for incorrect than for correct predictions in both regimes (Mann–Whitney p = 7.1 × 10⁻⁷⁷ with rank-biserial |r| = 0.30 under shift). In the laryngeal cohort, no statistically significant differences were detected among four architectures (ROC-AUC 0.844–0.901; all pairwise DeLong p > 0.05). Conclusions: Random-split evaluation substantially overestimated performance relative to a resolution-defined acquisition-shift stress test, and entropy-based selective prediction improved reliability by identifying a high-confidence subset for automated prediction while deferring the remainder to human review. Target-domain recalibration substantially restores calibration under shift (ECE 0.172 → 0.036 with temperature scaling; → 0.017 with isotonic regression) but does not recover discrimination; selective prediction is complementary, mitigating residual confident-wrong predictions. An encoder-transfer experiment showed asymmetric cross-domain utility; features learned on the larger gastrointestinal cohort transferred to the laryngeal cohort (AUROC 0.80 vs. in-domain 0.89), whereas the reverse direction did not transfer (0.53 vs. 0.72). Prospective multi-center validation remains required before clinical deployment. All code, fold definitions, random seeds, and a reproducible protocol are publicly released. Full article

Fiscal receipt text recognition is an important but difficult task within scene text recognition (STR). This study focuses on cropped word-level receipt text images whose text regions are already available, rather than on a complete end-to-end receipt OCR pipeline that also includes text detection and key information extraction. In real-world scenarios, receipt images often contain compact layouts, stamps, folds, uneven contrast, small fonts, and highly imbalanced character distributions. Rare symbols, including decimal points, slashes, colons, and currency marks, are especially important because recognition errors may directly affect financial information. Although Transformer-based recognizers such as PARSeq achieve strong performance on general STR tasks, their patch-token representation weakens local spatial priors, while the conventional cross-entropy loss gives insufficient emphasis to rare receipt-specific characters. To address these limitations, this paper presents CoordAtt-PARSeq, a lightweight extension of PARSeq-tiny for cropped fiscal receipt text recognition. The proposed model inserts a Coordinate Attention (CoordAtt) stem before patch embedding to strengthen horizontal and vertical spatial dependencies, and employs Focal Loss to improve learning on hard and under-represented characters. On WildReceipt, CoordAtt-PARSeq improves PARSeq-tiny by 4.19 percentage points in word accuracy (Ignore Case), 4.07 percentage points in strict word accuracy, and 1.11 percentage points in Normalized Edit Distance (NED), while reducing Character Error Rate (CER) by 1.85 absolute points. It also outperforms CRNN, ViTSTR, ABINet, TRBA, and PARSeq-tiny. On SROIE, the proposed method improves PARSeq-tiny by 4.47 percentage points in word accuracy (Ignore Case), 4.39 percentage points in strict word accuracy, and 1.50 percentage points in NED, with a 1.84-point absolute CER reduction. These results indicate that direction-aware spatial enhancement combined with imbalance-aware character learning can effectively adapt Transformer-based STR models to robust cropped fiscal receipt text recognition. Full article

Precision livestock farming requires monitoring approaches that extend beyond static activity thresholds to enable dynamic, animal-level decision support. A time-aware machine learning framework was developed to transform RFID-derived goat activity records into an interpretable behavioral monitoring system. Time-stamped activity data were processed into temporal features such as lagged activity, rolling mean, activity change, and elapsed-time variables to capture short-term behavioral history. The framework integrated latent-state discovery, fuzzy-uncertainty analysis, transition modeling, supervised classification, short-horizon forecasting, and dashboard-based alert visualization within a predictive dashboard-based monitoring framework. Four latent behavioral clusters were identified, with the dominant cluster representing a stable low-activity baseline and accounting for 77.88% of observations. Boundary-zone analysis indicated that 7.30% of observations were in transitional regions, while fuzzy clustering classified 21.42% as uncertain or mixed-state points, suggesting gradual shifts in activity. Transition analysis revealed greater persistence in baseline states and lower persistence in high-activity/non-baseline states, which exhibited the highest volatility and entropy. Using nested time-blocked validation, Random Forest predicted future high-activity/non-baseline onset with AUC values of 0.869 and 0.841 for 5 and 10 min horizons, respectively. These results demonstrate that activity instability can be detected and forecasted over short horizons, supporting behavior-based monitoring. However, external biological validation is still required before implementation as a health- or disease-detection system. Full article

Unsupervised domain adaptation (UDA) transfers knowledge from a labeled source domain to an unlabeled target domain and is particularly valuable in medical imaging, where dense annotations are costly and acquisition conditions vary. Cross-modality segmentation remains challenging because modality-dependent intensity and texture shifts alter image appearance, while teacher-generated pseudo-labels are often unreliable near anatomical boundaries. We propose a fractional texture-guided and boundary-aware perturbation-learning framework within a student–teacher scheme. On the source side, soft histogram transfer introduces target-related low-order intensity shifts. A multi-order fractional Gram discrepancy between shallow features of the intensity-transferred source and target images then provides a gradient signal for generating magnitude-normalized, range-clipped perturbations. This discrepancy is used as a perturbation cue rather than a direct alignment loss, exposing the student to target-relevant texture and edge-transition variation while preserving source annotations. On the target side, teacher logits are perturbed only within predicted boundary bands to model local contour uncertainty. Box-counting fractal boundary complexity guides the boundary-band width and logit perturbation scale and, together with predictive entropy, regulates pseudo-label supervision. Across five adaptation tasks, the proposed method achieves three-seed mean ± standard deviation Dice scores of 89.24 ± 0.12% and 82.01 ± 0.10% for cardiac MR→CT and CT→MR, 88.65 ± 0.29% and 90.43 ± 0.22% for abdominal MR→CT and CT→MR, and 84.76 ± 0.25% for bSSFP→LGE adaptation. Within the protocol-aware benchmark comparisons, the proposed method attains the highest average Dice score on four of the five tasks and is within 0.07 percentage points of the highest reported value on abdominal CT→MR. Ablation and operator-replacement studies further indicate that the source- and target-side pathways provide complementary benefits. Because all auxiliary perturbation and reliability-weighting modules are used only during adaptation, deployment requires only the adapted segmentation network, without additional inference-time modules or parameters. Full article