Neural Networks最新文献

D3I-COMCF: Dissimilarity-driven dual-interaction via co-occurrence motifs and complementary fragments for drug-drug interaction prediction. D3I-COMCF：通过共现基序和互补片段进行药物-药物相互作用预测的差异性驱动的双重相互作用。

IF 7.2 1区计算机科学

Neural Networks Pub Date : 2026-12-01 Epub Date: 2026-07-02 DOI: 10.1016/j.neunet.2026.109328

Wentao Xie, Min Jin, Mingjian Yang, Yulu Zhou, Xinhua Liu, Meng Li, Jagath C Rajapakse

{"title":"D3I-COMCF: Dissimilarity-driven dual-interaction via co-occurrence motifs and complementary fragments for drug-drug interaction prediction.","authors":"Wentao Xie, Min Jin, Mingjian Yang, Yulu Zhou, Xinhua Liu, Meng Li, Jagath C Rajapakse","doi":"10.1016/j.neunet.2026.109328","DOIUrl":"10.1016/j.neunet.2026.109328","url":null,"abstract":"Drug-drug interaction (DDI) prediction seeks to identify pharmacodynamic and pharmacokinetic interactions arising from the co-administration of multiple drugs with differing physicochemical properties. However, most existing methods predominantly focus on global molecular or atomic topologies, overlooking the critical role of dynamically derived functional substructures (motifs) in modulating drug activity and the occurrence and strength of DDI. To this end, we propose a dissimilarity-driven dual-interaction via co-occurrence motifs and complementary fragments (D3I-COMCF) framework that deeply integrates molecular-, motif-, and atomic-level information for accurate DDI prediction. Specifically, we construct, for the first time, a co-occurrence motifs molecular interaction graph (CMMIG) that unifies the global molecular structure, local physicochemical properties of motifs, and molecule-motif interaction topology. By weighting rare, dissimilar but information-rich motifs, D3I-COMCF was guided to focus on more discriminative functional units. Moreover, based on the maximum common substructure mapping and its dissimilar one-hop neighbor extension, we dynamically incorporated complementary fragments attached to the shared backbone into a fragment interaction graph, revealing electronic complementarity and spatial cooperativity with computational efficiency. Finally, hierarchical multi-block integration of the three levels enables cross-scale joint representation of drug pairs. Extensive experiments on multiple public datasets verified the effectiveness of D3I-COMCF, which achieved strong and stable performance under standard benchmark protocols and showed promising cold-start generalization and interpretability. Additional hard-negative analysis further indicates that D3I-COMCF remains sensitive to known positive DDIs, while controlling false positives among structurally similar candidate pairs not annotated as known positives remains challenging.","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"204 ","pages":"109328"},"PeriodicalIF":7.2,"publicationDate":"2026-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148406673","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

A boundary-regularization-enhanced video anomaly detection network based on context-adaptive spatio-temporal conditional diffusion. 基于上下文自适应时空条件扩散的边界正则化增强视频异常检测网络。

IF 7.2 1区计算机科学

Neural Networks Pub Date : 2026-12-01 Epub Date: 2026-07-02 DOI: 10.1016/j.neunet.2026.109330

Huilin Liu, Guanghan Sun, Xiaolong Hu, Tian Liu, Wanqi Ma

{"title":"A boundary-regularization-enhanced video anomaly detection network based on context-adaptive spatio-temporal conditional diffusion.","authors":"Huilin Liu, Guanghan Sun, Xiaolong Hu, Tian Liu, Wanqi Ma","doi":"10.1016/j.neunet.2026.109330","DOIUrl":"10.1016/j.neunet.2026.109330","url":null,"abstract":"Unsupervised video anomaly detection aims to learn normal spatio-temporal patterns from unlabeled surveillance data and identify deviations during inference. Existing methods often struggle to capture long-term temporal dependencies and adapt to spatial variations caused by target deformation and viewpoint changes. Although diffusion-based models provide stable distribution modeling, their limited temporal conditioning and fixed spatial representations restrict anomaly discrimination in complex scenes. To address these issues, we propose a boundary regularization enhanced video anomaly detection network based on context-adaptive spatio-temporal conditional diffusion. Rather than relying on rigid feature extraction, our method explicitly models long-range motion evolution and dynamically adapts to spatial variations caused by target deformation and viewpoint shifts. Furthermore, to prevent overfitting to normal behaviors, we introduce a hybrid pseudo-anomaly generation strategy that acts as a boundary regularization constraint, explicitly sharpening the model's discriminative capability against abnormal patterns. Experiments on benchmark datasets including UCSD Ped2, CUHK Avenue, and ShanghaiTech demonstrate that our method significantly improves detection accuracy. By effectively integrating context-adaptive conditioning and boundary regularization, this work establishes a highly discriminative and generalizable paradigm for video anomaly detection, demonstrating strong robustness in complex, real-world surveillance scenarios.","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"204 ","pages":"109330"},"PeriodicalIF":7.2,"publicationDate":"2026-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148399058","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

DA-S²RL: Disease-aware shared-specific representation learning for heterogeneous image-EHR Alzheimer's disease diagnosis.

IF 7.2 1区计算机科学

Neural Networks Pub Date : 2026-10-19 DOI: 10.1016/j.neunet.2026.109760

Wei Wang, Yu Shi, Juan Cheng, Yu Liu

{"title":"DA-S2RL: Disease-aware shared-specific representation learning for heterogeneous image-EHR Alzheimer's disease diagnosis.","authors":"Wei Wang, Yu Shi, Juan Cheng, Yu Liu","doi":"10.1016/j.neunet.2026.109760","DOIUrl":"https://doi.org/10.1016/j.neunet.2026.109760","url":null,"abstract":"Combining electronic health records (EHR) with medical images is important for improving the diagnosis of Alzheimer's disease (AD). However, neuroimages and EHR differ greatly in semantic content and data distribution. Neuroimages mainly reflect brain structure and metabolic abnormalities, while EHR records clinical symptoms and behavioral changes. Directly fusing these heterogeneous modalities may mix shared disease information with modality-specific diagnostic cues, resulting in redundant and ambiguous representations. To address this problem, we propose DA-S2RL, a disease-aware shared-specific multimodal representation learning framework for AD diagnosis. DA-S2RL organizes heterogeneous image and EHR features into disease-related shared semantics and modality-specific diagnostic residuals. The core of DA-S2RL is the shared-specific cross-modal coding (SCC) module. It strengthens cross-modal semantic consistency, suppresses redundant correlations, and enables mutual refinement between shared semantics and diagnostic residuals. Before SCC, image features are integrated into a compact image representation, and EHR features are encoded into the same feature space. Experiments on ADNI and NACC demonstrate that explicitly organizing heterogeneous image-EHR representations into shared disease semantics and modality-specific residuals improves multimodal diagnostic learning under different image-EHR settings.","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"206 Pt A","pages":"109760"},"PeriodicalIF":7.2,"publicationDate":"2026-10-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149225413","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

AGformer: Adaptive filtering and gated fusion-attention for multivariate time series forecasting.

IF 7.2 1区计算机科学

Neural Networks Pub Date : 2026-10-19 DOI: 10.1016/j.neunet.2026.109759

Shu-Xin Xu, Jiang-Wen Xiao, Peng-Cheng Li, Tian-Hao Lu

{"title":"AGformer: Adaptive filtering and gated fusion-attention for multivariate time series forecasting.","authors":"Shu-Xin Xu, Jiang-Wen Xiao, Peng-Cheng Li, Tian-Hao Lu","doi":"10.1016/j.neunet.2026.109759","DOIUrl":"https://doi.org/10.1016/j.neunet.2026.109759","url":null,"abstract":"Multivariate time series forecasting is a fundamental problem in sequence modeling, with broad applications spanning energy systems, meteorology, finance, and transportation. Despite significant progress in Transformer-based, TCN-based, and MLP-based architectures, three persistent neural network design challenges remain unresolved: 1) non-adaptive frequency-domain filtering that applies identical spectral parameters across heterogeneous variables; 2) inadequate utilization of informative short-term fluctuations that are discarded as noise yet encode critical event-driven dynamics; and 3) insufficient exploration of inter-scale dependencies in multi-scale temporal representations. To address these challenges, this paper proposes AGformer, a novel forecasting architecture integrating three complementary neural modules: 1) A variable-oriented Adaptive Multi-k Filter (AKF) that dynamically generates per-variable spectral filtering parameters from learned frequency-domain features, eliminating the need for manual global hyperparameter tuning. 2) A Dual-Gate Fusion-Attention (DGFA) mechanism that synergistically models long-term trends and short-term fluctuations via complementary temporal and content gating, recovering event-driven signals that are otherwise discarded by standard denoising. 3) A Multi-Scale Embedding (MSED) layer that fuses temporal patterns across resolutions through attention-based inter-scale weighting, capturing cross-resolution hierarchical dependencies that simple concatenation or summation cannot model. Extensive experiments on nine representative benchmark datasets across electricity, weather, finance, and transportation domains demonstrate AGformer's consistent superiority, achieving average reductions of 8.1%-23.6% in MSE and MAE relative to state-of-the-art baselines across nine benchmarks.","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"206 Pt A","pages":"109759"},"PeriodicalIF":7.2,"publicationDate":"2026-10-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149225407","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Dynamic gain neuromodulation attenuates stability gaps during continual joint training.

IF 7.2 1区计算机科学

Neural Networks Pub Date : 2026-10-14 DOI: 10.1016/j.neunet.2026.109767

Alejandro Rodriguez-Garcia, Anindya Ghosh, Srikanth Ramaswamy

{"title":"Dynamic gain neuromodulation attenuates stability gaps during continual joint training.","authors":"Alejandro Rodriguez-Garcia, Anindya Ghosh, Srikanth Ramaswamy","doi":"10.1016/j.neunet.2026.109767","DOIUrl":"https://doi.org/10.1016/j.neunet.2026.109767","url":null,"abstract":"Recent work in continual learning has highlighted the stability gap - a temporary performance drop on previously learned tasks when new ones are introduced. This phenomenon reflects a mismatch between rapid adaptation and strong retention at task boundaries, underscoring the need for optimization mechanisms that balance plasticity and stability over abrupt distribution changes. While optimizers such as momentum-SGD and Adam introduce implicit multi-timescale behavior, they still exhibit pronounced stability gaps. Importantly, these gaps persist even under ideal joint training, making it crucial to study them in this setting to isolate their causes from other sources of forgetting. Motivated by how noradrenergic (neuromodulatory) bursts transiently increase neuronal gain under uncertainty, we introduce a dynamic gain scaling mechanism as a two-timescale optimization technique that balances adaptation and retention by transiently increasing the effective update magnitude while dynamically reparameterizing the weights governing the forward pass, thereby empirically mitigating transition-induced curvature amplification. Across domain- and class-incremental MNIST, CIFAR, and mini-ImageNet benchmarks under task-agnostic joint training, dynamic gain scaling effectively attenuates stability gaps while maintaining competitive accuracy, improving robustness at task transitions.","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"206 Pt A","pages":"109767"},"PeriodicalIF":7.2,"publicationDate":"2026-10-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149225395","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

A single-loop bilevel deep learning method for optimal control of obstacle problems.

IF 7.2 1区计算机科学

Neural Networks Pub Date : 2026-10-14 DOI: 10.1016/j.neunet.2026.109761

Yongcun Song, Shangzhi Zeng, Jin Zhang, Lvgang Zhang

{"title":"A single-loop bilevel deep learning method for optimal control of obstacle problems.","authors":"Yongcun Song, Shangzhi Zeng, Jin Zhang, Lvgang Zhang","doi":"10.1016/j.neunet.2026.109761","DOIUrl":"https://doi.org/10.1016/j.neunet.2026.109761","url":null,"abstract":"Optimal control of obstacle problems arises in a wide range of applications and is computationally challenging due to the nonsmoothness, nonlinearity, and bilevel structure of the underlying problems. Classical numerical methods typically rely on mesh-based discretizations and require solving a sequence of computationally expensive subproblems, while deep learning approaches for this class of problems remain largely unexplored. In this work, we propose a bilevel deep learning framework whose core training algorithm is the computationally efficient Single-Loop Stochastic First-Order Bilevel Algorithm (S2-FOBA), which avoids nested optimization, uses only first-order information, and does not rely on restrictive lower-level uniqueness assumptions. The proposed method is mesh-free, scalable to high-dimensional settings and complex domains, and avoids the repeated solution of discretized subproblems. It employs constraint-embedding neural networks to approximate both the state and the control, thereby enforcing admissibility by construction while preserving the intrinsic bilevel structure of the problem. We further analyze the convergence behavior of S2-FOBA under mild conditions. Numerical experiments on benchmark problems, including two-dimensional examples with regular and irregular obstacles, a five-dimensional example, and a problem posed on a complex domain, demonstrate the effectiveness, efficiency, and robust constraint enforcement of the proposed approach.","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"206 Pt A","pages":"109761"},"PeriodicalIF":7.2,"publicationDate":"2026-10-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149225358","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

MIMOT: Mamba-based inter-trajectory interaction for multi-object tracking.

IF 7.2 1区计算机科学

Neural Networks Pub Date : 2026-10-14 DOI: 10.1016/j.neunet.2026.109766

Biao Liu, Yanxin Wu, Jiaqi Yang, Junchao Qiao

{"title":"MIMOT: Mamba-based inter-trajectory interaction for multi-object tracking.","authors":"Biao Liu, Yanxin Wu, Jiaqi Yang, Junchao Qiao","doi":"10.1016/j.neunet.2026.109766","DOIUrl":"https://doi.org/10.1016/j.neunet.2026.109766","url":null,"abstract":"Online multi-object tracking (MOT) requires reliable motion prediction when detections are missing or targets occlude one another. Existing drop-in state-space model (SSM) motion predictors mainly derive each update from the target's own history, which limits their use of nearby motion context. We propose MIMOT, an online tracking-by-detection method that replaces the Kalman filter with a geometry-aware Mamba motion predictor while retaining YOLOX detection and ByteTrack-style association without appearance cues. MIMOT first encodes contiguous and temporally strided track histories and fuses their complementary temporal evidence. It then conditions nearby trajectories on relative geometry and aggregates them through a distance-ordered bidirectional Mamba scan whose aggregation cost is linear in the selected neighborhood size. A codebook-conditioned SSM decoder subsequently predicts residual box updates using input-dependent mixtures of transition parameters. MIMOT achieves 58.1, 74.2, and 62.1 HOTA on the official DanceTrack, SportsMOT, and MOT17 test sets, respectively. It ranks first among the evaluated learned-motion trackers on DanceTrack and MOT17 and remains competitive on SportsMOT. Controlled ablations further show that temporal fusion, geometry-conditioned interaction, and codebook decoding mainly improve association accuracy.","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"206 Pt A","pages":"109766"},"PeriodicalIF":7.2,"publicationDate":"2026-10-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149224865","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Lagrangian-based safe Q-learning for nonlinear delayed stochastic systems.

IF 7.2 1区计算机科学

Neural Networks Pub Date : 2026-10-13 DOI: 10.1016/j.neunet.2026.109752

Kaidi Xing, Zhongyang Ming, Jiayue Sun

{"title":"Lagrangian-based safe Q-learning for nonlinear delayed stochastic systems.","authors":"Kaidi Xing, Zhongyang Ming, Jiayue Sun","doi":"10.1016/j.neunet.2026.109752","DOIUrl":"https://doi.org/10.1016/j.neunet.2026.109752","url":null,"abstract":"This paper addresses the safe optimal control problem for a class of continuous-time nonlinear Itô stochastic systems with state delays and unknown dynamics. We propose a safe Q-learning algorithm that, for the first time, extends this framework from linear deterministic systems to the nonlinear stochastic domain. The simultaneous presence of stochastic perturbations and state delays poses substantial challenges to safety enforcement and optimal policy learning. To address these challenges, a barrier Lyapunov function (BLF) and a Lagrangian saddle-point formulation are employed to characterize the safety-constrained optimal control problem. An adaptive safety multiplier estimation mechanism is then extended to stochastic systems with state delays, yielding a state- and delay-dependent safety mechanism that dynamically adjusts the intervention intensity and thereby alleviates the conservatism associated with fixed-gain safety intervention. Furthermore, a structured safe Q-function approximator and a model-free integral learning law are developed to avoid online system identification, whose computational burden and approximation errors can be exacerbated by state delays and stochastic perturbations. A Lyapunov-Krasovskii-based analysis establishes stochastic safety by deriving a uniform bound on the expected BLF together with a probabilistic compact-subset guarantee. It is further shown that the closed-loop state and the Critic weight estimation error are uniformly ultimately bounded in the mean-square sense.","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"206 Pt A","pages":"109752"},"PeriodicalIF":7.2,"publicationDate":"2026-10-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149225400","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

A spatio-temporal attention convolutional network with topology-preserving channel embedding and transformer for EEG decoding.

IF 7.2 1区计算机科学

Neural Networks Pub Date : 2026-10-12 DOI: 10.1016/j.neunet.2026.109755

Wenqi Sun, Shuo Fu, Jinhua Zhang, Aibin Zhu, Jing Wang, Shuai Yin, Gang Liu

{"title":"A spatio-temporal attention convolutional network with topology-preserving channel embedding and transformer for EEG decoding.","authors":"Wenqi Sun, Shuo Fu, Jinhua Zhang, Aibin Zhu, Jing Wang, Shuai Yin, Gang Liu","doi":"10.1016/j.neunet.2026.109755","DOIUrl":"https://doi.org/10.1016/j.neunet.2026.109755","url":null,"abstract":"EEG-based BCI has shown promising applications in motor rehabilitation and emotion recognition. Given the non-stationarity and variability of EEG signals, it is essential to develop classification models that can effectively extract spatio-temporal features while maintaining strong generalization capabilities. This paper proposes a spatio-temporal attention convolutional network for EEG decoding (E-STACNet) to encode the topological features and temporal characteristics of EEG signals within a unified network framework. It consists of channel-attention (CA) block, convolutional (CV) layers, multi-head self-attention (MSA) and classifier layer. Specifically, CA block encodes spatial topological feature via a topology-preserving channel embedding method. Then convolutional layers are employed to encode low-level temporal information and further extract spatial features, followed by MSA block to highlight the most important temporal information of EEG signals. Finally, the outputs are obtained through two fully-connected layers. We conducted experiments on three public EEG-based datasets including BCI Competition IV dataset 2a, Giga DB and DEAP. The results demonstrate E-STACNet achieves average accuracies of 88.00% and 80.62% on datasets 2a and Giga DB, respectively. On the DEAP, it reaches average accuracies of 96.49% and 96.46% for valence and arousal. The proposed model outperforms other state-of-the-art models on all three datasets. Additionally, the ablation study reveals the importance of topology-preserving channel embedding for the proposed network. Visualization analysis shows E-STACNet could obtain the importance of channels at different locations through the channel-attention mechanism, which demonstrates the interpretability of the proposed model. Overall, the accurate and generalization capabilities of E-STACNet will promote the development of EEG-based BCIs.","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"206 Pt A","pages":"109755"},"PeriodicalIF":7.2,"publicationDate":"2026-10-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149225351","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Bidirectional cross-frequency guided wavelet diffusion for detail-preserving low-light image enhancement. 双向交叉频导小波扩散保细节弱光图像增强。

IF 7.2 1区计算机科学

Neural Networks Pub Date : 2026-09-19 DOI: 10.1016/j.neunet.2026.109650

Pingping Liu, Jiahao Liu, Shihui Pei, Tongshun Zhang, Qingyi Xue, Kai Zhang, Qiuzhan Zhou

{"title":"Bidirectional cross-frequency guided wavelet diffusion for detail-preserving low-light image enhancement.","authors":"Pingping Liu, Jiahao Liu, Shihui Pei, Tongshun Zhang, Qingyi Xue, Kai Zhang, Qiuzhan Zhou","doi":"10.1016/j.neunet.2026.109650","DOIUrl":"https://doi.org/10.1016/j.neunet.2026.109650","url":null,"abstract":"Low-light image enhancement is challenging due to noise and poor visibility. While diffusion models have shown remarkable generative capabilities, applying them directly to this task in the pixel domain remains problematic. They tend to struggle with the trade-off between global illumination correction and local detail preservation, often resulting in unnatural textures or amplified noise due to the lack of frequency-specific constraints. To address these limitations, we propose a novel Bidirectional Cross-Frequency Guided Wavelet Diffusion model (BCF-Diff), which operates in the frequency domain for efficient and detail-preserving enhancement. Our approach uses K-level Discrete Wavelet Transform to decompose images. We introduce a two-pronged strategy to tackle the core challenges. First, we focus the diffusion model exclusively on restoring the compact low-frequency component for efficiency and stability with a detail-weighted loss mechanism. This mechanism computes weights from the high-frequency components to encourage the model to prioritize the reconstruction of detail-rich regions, ensuring a high-quality structural base. Then, we propose an innovative Cross-Frequency Gated Refinement (CFGR), which leverages the newly restored low-frequency component as a guidance to direct the enhancement of high-frequency details, and subsequently employs a Spatial Gated Refinement Module to adaptively suppress noise and sharpen textures. Our bidirectional cross-frequency guidance ensures that textures and edges are generated coherently with the overall scene structure. Experiments show that BCF-Diff achieves competitive overall performance and particularly strong structural similarity on the evaluated paired datasets, while providing visually detailed enhancement results under low-light conditions.","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"205 Pt C","pages":"109650"},"PeriodicalIF":7.2,"publicationDate":"2026-09-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149026570","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0