{"title":"MKDFusion: modality knowledge decoupled for infrared and visible image fusion","authors":"Yucheng Zhang, You Ma, Lin Chai","doi":"10.1007/s10489-025-06470-w","DOIUrl":"10.1007/s10489-025-06470-w","url":null,"abstract":"<div><p>The purpose of infrared and visible fusion is to integrate useful information from both infrared and visible images into a single image. The fused image should possess rich texture details and salient target information of the two images. Current image fusion algorithms primarily face two limitations: 1) The lack of decoupling between modality-agnostic and modality-specific knowledge during the feature extraction stage hinders the alignment of modality-agnostic knowledge and the differentiation of modality-specific knowledge. 2) The interaction between modality features is not sufficiently explored in the feature fusion stage, which inhibits the exploitation of complementary information. To address the above challenges, we propose a Modality Knowledge Decoupled (MKD) module in the feature extraction stage and a Cross-Modality Mamba Fusion (CMF) module in the feature fusion stage. In MKD, we first utilize a dual-branch network to extract modality-agnostic and modality-specific knowledge separately. Then, a pair of Knowledge Discriminators (KD) is constructed to minimize inter-modality irrelevant knowledge and maximize inter-modality relevant knowledge. In CMF, the interactions between different modality knowledge are learnt in a hidden state space, which not only reduces the inter-modality knowledge differences but also enhances the texture information of the image. Experiments on three datasets demonstrate that our method outperforms existing methods, highlighting less salient targets and texture information more effectively. In addition, MKDFusion has demonstrated excellent generalization performance and enormous potential in high-level vision tasks in medical image fusion and object detection applications. The code is available at https://github.com/SEU-ZYC/MKDFusion.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"55 7","pages":""},"PeriodicalIF":3.4,"publicationDate":"2025-04-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143826570","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Qingwei Sun, Jiangang Chao, Wanhong Lin, Wei Chen, Zhenying Xu, Jin Yang
{"title":"Few-shot segmentation combined with domain adaptation: a flexible paradigm for parsing astronaut work environments","authors":"Qingwei Sun, Jiangang Chao, Wanhong Lin, Wei Chen, Zhenying Xu, Jin Yang","doi":"10.1007/s10489-025-06508-z","DOIUrl":"10.1007/s10489-025-06508-z","url":null,"abstract":"<div><p>The capacity to perform few-shot segmentation of the astronaut work environment (AWE) is of critical importance, especially for tasks that cannot be predetermined. The challenging task of transferring FSS models, which are trained on natural datasets, to the AWE—referred to as cross-domain few-shot segmentation (CD-FSS)—holds substantial importance. Rather than devising an entirely novel model, we propose an approach that integrate domain adaptation (DA) with extant FSS models, herein termed meta learners. Specifically, a prior learner based on generative adversarial networks (GAN) is devised to impart semantic guidance to the meta learner. To discern challenging samples, a loss function incorporating a scaling factor is employed during the training stage of the prior learner. Furthermore, a metric-based fusion module is proposed to mitigate bias in accordance with the association between the prior learner and the meta learner. The results evince that our method can be seamlessly integrated with different types of existing FSS models, thereby enhancing their cross-domain performance.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"55 7","pages":""},"PeriodicalIF":3.4,"publicationDate":"2025-04-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143830679","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"MGMP: Multi-granularity semantic relation learning and meta-path structure interaction learning for fake news detection","authors":"Baozhen Lee, Dandan Cao, Tingting Zhang","doi":"10.1007/s10489-025-06560-9","DOIUrl":"10.1007/s10489-025-06560-9","url":null,"abstract":"<div><p>This paper proposes the joint learning model <u>M</u>ulti-<u>G</u>ranularity Semantic Relation Learning and <u>M</u>eta-<u>P</u>ath Structure Interaction Learning for fake news detection (MGMP). The MGMP improves global semantic relation learning through a multi-granularity process involving coarse-grained and fine-grained learning modules, along with meta-path based global interaction learning. It begins by refining global semantic recognition accuracy at the word-level and document-level through attention mechanisms and convolutional neural networks. Furthermore, it enhances global interaction learning by enhancing meta-path instance representations with various meta-paths and employing multi-head self-attention mechanisms within the network structure. Experimental findings on real datasets confirm the effectiveness of the MGMP in fake news detection by enhancing global semantic recognition accuracy in news nodes and recognizing network structural characteristics.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"55 7","pages":""},"PeriodicalIF":3.4,"publicationDate":"2025-04-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143826602","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"Performance-based active learning (PbAL) for imbalanced data with nonparametric logistic regression","authors":"Wonjae Lee, Kangwon Seo","doi":"10.1007/s10489-025-06531-0","DOIUrl":"10.1007/s10489-025-06531-0","url":null,"abstract":"<div><p>Real-world data often exhibit asymmetric class distributions, where certain target values have significantly fewer observations compared to the others. This lack of uniform distribution across categories can substantially affect model performance in classification problems. This research introduces the performance-based active learning (PbAL) scheme to address the class imbalance problem considering the nonlinear decision boundary. PbAL is designed to sequentially select the most beneficial samples from an imbalanced data set by directly evaluating a performance metric on a pool of data. While parametric logistic regression offers a fundamental classification model with ease of interpretation, the assumption of linear relationship in the logit function is often questionable. The use of nonparametric logistic regression with smoothing splines allows for a more flexible classification boundary. Experiments with several data sets demonstrate that PbAL often outperforms traditional active learning approaches based on D-optimality and A-optimality. Additionally, the proposed method yields superior results compared to other resampling techniques commonly used for imbalanced classification problems even with a smaller sample size. These findings suggest that PbAL effectively mitigates bias caused by training on imbalanced classes, which can severely impact model’s ability to accurately predict class labels for new observations.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"55 7","pages":""},"PeriodicalIF":3.4,"publicationDate":"2025-04-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143818295","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"Trajectory optimization of train cooperative energy-saving operation using a safe deep reinforcement learning approach","authors":"Wenguang Niu, Yonghua Zhou, Xiangmeng Jiao, Hamido Fujita, Hanan Aljuaid","doi":"10.1007/s10489-025-06542-x","DOIUrl":"10.1007/s10489-025-06542-x","url":null,"abstract":"<div><p>Energy-efficient optimization of train speed profiles can effectively reduce the traction energy consumption of urban rail transit systems. Existing reinforcement learning (RL) optimization models for optimizing train operation profiles do not proactively handle the utilization constraints of regenerative braking energy (RBE). For this reason, this paper proposes an optimization model of train energy-saving profiles under multi-train cooperative operations. A novel safe deep reinforcement learning algorithm, guided by heuristic rules, is developed to optimize energy-saving train driving strategies in various scenarios. To ensure safety during the agent’s learning processes, a two-layer protection mechanism with soft constraint and truncation penalties is employed. Dynamic energy constraints are also introduced to enable the RBE utilization between trains. The simulation experiments using a real metro line data show that the proposed model and algorithm not only generate safe and energy-efficient profiles that meet metro operational constraints but also maximize the RBE utilization between trains, significantly reducing traction energy consumption.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"55 7","pages":""},"PeriodicalIF":3.4,"publicationDate":"2025-04-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143818296","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Jian Cen, Weiwei Si, Xi Liu, Bichuang Zhao, Hankun Huang, Junfu Liu
{"title":"Intelligent fault diagnosis method based on data generation and long-patch vision transformer under small samples","authors":"Jian Cen, Weiwei Si, Xi Liu, Bichuang Zhao, Hankun Huang, Junfu Liu","doi":"10.1007/s10489-025-06535-w","DOIUrl":"10.1007/s10489-025-06535-w","url":null,"abstract":"<div><p>Rotating machinery is an important part of modern industry, and bearings are one of the most important things. However, bearing fault data are difficult to collect, and bearing fault diagnosis under small samples has significant research potential. In this paper, we proposed a fault diagnosis framework that combines diffusion modeling and improved Vision Transformer. First, the short-time Fourier transform is applied to the original one-dimensional vibration signals to convert the data into time-frequency maps. Second, the conditional diffusion model was applied to generate the required samples and expand the dataset. Finally, the Long-patch Vision Transformer (LVT) proposed in this paper is used to classify the mixed samples. LVT designs a long-patch division method for time-frequency maps with dense transverse features. The LVT contains denser features in each patch, and this method is more suitable for time-frequency maps. Validating the method proposed in this paper on two datasets and comparing it with other methods, our method achieved the highest accuracy among the compared methods.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"55 7","pages":""},"PeriodicalIF":3.4,"publicationDate":"2025-04-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143818294","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"TRSD: tensor spatial reconstruction and spectral metric decision fusion for hyperspectral anomaly detection with noise","authors":"Zhenhua Mu, Yihan Wang, Xianghai Wang","doi":"10.1007/s10489-025-06504-3","DOIUrl":"10.1007/s10489-025-06504-3","url":null,"abstract":"<div><p>The unique and detailed spectral information in hyperspectral images (HSI) provides an advantage for distinguishing different targets in anomaly detection (AD). However, most traditional HSI-AD methods primarily focus on the inherent spectral structure information, often overlooking the strong spatial-spectral synergy present in HSI. An increase in spectral resolution typically leads to a decrease in the number of photons received per channel, which increases the likelihood of correlated noise during image formation. To address these issues and significantly improve detection performance, a method called Tensor Space Reconstruction and Spectral Local Correlation Metric Decision Fusion (TRSD) is proposed for HSI-AD in the presence of noise. First, three-dimensional principal component (PC) extraction, based on information entropy, is performed to obtain a denoised purified image for reconstruction. The initial feature detection image is generated by calculating the purified image using the local Mahalanobis distance. To compensate for the loss of spectral information caused by PC analysis in the spectral dimension during Tucker reconstruction, the feature map is extracted using the local spectral correlation metric. Finally, the two detection feature images are adaptively fused to generate the final AD image, which highlights anomaly targets and improves detection accuracy.The proposed algorithm is experimentally validated through comparisons with current typical AD algorithms, using real HSIs captured in four different complex noise-added scenarios. The effectiveness of the algorithm is demonstrated through experiments. The source code for TRSD will be made publicly available at https://github.com/muzhenhuam/TRSD.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"55 7","pages":""},"PeriodicalIF":3.4,"publicationDate":"2025-04-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143809162","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Anna Sztyber-Betley, Elodie Chanthery, Louise Travé-Massuyès, Gustavo Pérez-Zuñiga
{"title":"Diagnosis test selection for distributed systems under communication and privacy constraints","authors":"Anna Sztyber-Betley, Elodie Chanthery, Louise Travé-Massuyès, Gustavo Pérez-Zuñiga","doi":"10.1007/s10489-025-06543-w","DOIUrl":"10.1007/s10489-025-06543-w","url":null,"abstract":"<div><p>Distribution is often necessary for large-scale systems because it makes monitoring and diagnosis more manageable from both computational and communication costs perspectives. Decomposing the system into subsystems may also be required to satisfy geographic, functional, or privacy constraints. The selection of diagnosis tests guaranteeing some level of diagnosability must adhere to this decomposition by remaining as local as possible in terms of the required sensor variables. This helps minimize communication costs. In practical terms, this means that the number of interconnections between subsystems should be minimized while keeping diagnosability, i.e., fault isolation capability, at its maximum. This paper differentiates itself from existing literature by leveraging flexibility in forming the subsystems. Through structural analysis and graph partitioning, we address the combined challenges of constrained decomposition of a large-scale system into subsystems and the selection of diagnosis tests that achieve maximal diagnosability with minimal subsystem interconnection. The proposed solution is implemented through an iterative algorithm, which is proven to converge. Its efficiency is demonstrated using a case study in the domain of water networks.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"55 7","pages":""},"PeriodicalIF":3.4,"publicationDate":"2025-04-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143809173","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Peidong Li, Zhenghong Zhong, Yangguang Zhao, Changheng Shao, Yi Sui, Rencheng Sun
{"title":"Power-GNN: a graph over-sampling method to mitigate power-law distribution in graph neural networks","authors":"Peidong Li, Zhenghong Zhong, Yangguang Zhao, Changheng Shao, Yi Sui, Rencheng Sun","doi":"10.1007/s10489-025-06421-5","DOIUrl":"10.1007/s10489-025-06421-5","url":null,"abstract":"<div><p>Since the advent of Graph Neural Networks (GNNs), they have been widely applied in the analysis and processing of graph data, especially demonstrating outstanding performance in semi-supervised node classification tasks. However, the class distribution in real-world graph data often exhibits a long-tail, imbalanced distribution, posing significant challenges to the classification performance of GNNs. Graph over-sampling methods address this by synthesizing new nodes for minority classes and creating corresponding edges, thus aiming to balance class representation and enhance model accuracy. Nonetheless, the degree distribution of nodes in reality also follows a power-law distribution, leading to synthesized nodes becoming low-degree tail nodes under existing edge construction strategies. This restricts their ability to acquire sufficient aggregation information, thereby degrading their representation quality and impacting classification outcomes. To address these challenges, this paper introduces Power-GNN, a novel graph data over-sampling framework tailored to tackle the dual challenges of imbalanced class distribution and the power-law distribution of node degrees. Power-GNN innovatively utilizes the power-law distribution of node degrees in a reverse manner. It strategically adds edges with high similarity to nodes with fewer connections, thereby amplifying the aggregation capability of synthesized nodes and boosting overall model performance. Through evaluations on multiple public benchmark datasets, Power-GNN has demonstrated superior performance over existing baselines across three common GNN architectures.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"55 7","pages":""},"PeriodicalIF":3.4,"publicationDate":"2025-04-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143809240","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Chengcheng Zhong, Kai Zhang, Zitong Zhang, Yanan Jiang, Chunlei Zhang
{"title":"DF(^2)Net: deformable fourier filter network for hyperspectral image classification","authors":"Chengcheng Zhong, Kai Zhang, Zitong Zhang, Yanan Jiang, Chunlei Zhang","doi":"10.1007/s10489-025-06493-3","DOIUrl":"10.1007/s10489-025-06493-3","url":null,"abstract":"<div><p>MLP-like architectures in hyperspectral image (HSI) classification flourish recently. However, these methods face challenges such as insufficient spectral-spatial feature extraction capability and excessive consumption of network computing resources. To address these problems, a deformable Fourier filter network (DF<span>(^{varvec{2}})</span>Net) is proposed as an innovative lightweight MLP framework for HSI classification. DF<span>(^{varvec{2}})</span>Net employs Fourier transform filters and spatial deformable operations to efficiently capture spectral-spatial features while maintaining a lightweight design. Specifically, two modules in DF<span>(^{varvec{2}})</span>Net are developed to extract and facilitate the deep integration of spectral-spatial features, namely the spectral discrete Fourier transform filter (SeDFT) module and the spatial deformable discrete Fourier transform filter (SaD<span>(^{varvec{2}})</span>FT) module. The SeDFT module employs a one-dimensional discrete Fourier transform filter (1D<span>(^{varvec{2}})</span>FT) to extract spectral features in the frequency domain, effectively capturing detailed information from the original spectrum. Additionally, the parameter-free design of the SeDFT module streamlines the feature processing pipeline and improves computational efficiency. The SaD<span>(^{varvec{2}})</span>FT module performs a two-dimensional deformable discrete Fourier transform (2D<span>(^{varvec{3}})</span>FT) filter, enabling low-parameter feature extraction by transforming spatial features into frequency domain representations. Moreover, the spatial deformable operation enhances the capacity of the network to perceive spatial structural variations by introducing learnable offsets. Experimental results on four public HSI datasets demonstrate that DF<span>(^{varvec{2}})</span>Net consistently achieves superior performance in lightweight classification. Compared to other state-of-the-art models, DF<span>(^{varvec{2}})</span>Net significantly reduces both the number of parameters and computational resource requirements while preserving high performance.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"55 7","pages":""},"PeriodicalIF":3.4,"publicationDate":"2025-04-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143809161","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}