During the past decade, representation-based classification methods have received considerable attention in pattern recognition. In particular, the recently proposed non-negative representation based classification (NRC) method has been reported to achieve promising results in a wide range of classification tasks. However, NRC has two major drawbacks. First, there is no regularization term in the formulation of NRC, which may result in unstable solution and misclassification. Second, NRC ignores the fact that data usually lies in a union of multiple affine subspaces, rather than linear subspaces in practical applications. To address the above issues, this paper presents an affine non-negative collaborative representation (ANCR) model for pattern classification. To be more specific, ANCR imposes a regularization term on the coding vector. Moreover, since each point in an affine subspace can be expressed as an affine combination of other points in this affine subspace, ANCR introduces an affine constraint to better represent the data from affine subspaces. The experimental results on several benchmarking datasets demonstrate the merits of the proposed ANCR method. Concretely, on the Hopkins and Aircraft datasets, ANCR achieves accuracy of 97.8% and 87.7%, respectively, which represents an improvement of 2.2% and 0.4% over NRC. The source code of our ANCR is publicly available at https://github.com/yinhefeng/ANCR.

Predictive maintenance (PdM) based on Remaining Useful Life (RUL) prediction plays a crucial role in improving performance and reducing lifecycle costs of industrial equipment. This study proposes an intelligent PdM framework that integrates a RUL prediction model based on probabilistic neural network with a distributional reinforcement learning agent based on QR-DQN. In the first stage, the RUL prediction model is developed to process sensor data to generate accurate RUL predictions, quantify predictive uncertainty, and estimate the probability of failure within a given horizon. Building on the health condition assessment, the QR-DQN agent learns the distribution of long-term maintenance returns and makes sequential decisions among multiple actions. By adopting risk-sensitive decision rules, the agent explicitly accounts for uncertainty and failure risk, achieving a balance between safety, cost efficiency, and timeliness of interventions. Experimental evaluations on complex system degradation demonstrate that the proposed intelligent PdM outperforms conventional baselines by reducing catastrophic failures, optimizing maintenance schedules, and improving overall reliability.

In recent years, Dempster-Shafer evidence theory has been widely applied in multi-source information fusion. To address the unreasonable results under highly conflicting evidence, many methods have been proposed, particularly uncertainty-based weighting methods. However, these methods exhibit a negative weighting phenomenon in certain scenarios, where abnormal evidence is assigned higher weight than normal evidence. This paper proposes a multi-uncertainty clustering method by systematically analyzing the limitations of uncertainty-based weighting methods. We employ Spearman correlation coefficients to select appropriate uncertainty measures. These selected measures are calculated for evidence sources and then input into an improved K-Means algorithm for evidence clustering. For each formed evidence cluster, average evidence is generated to enhance the expression of intra-cluster common features. The support degree of different categories is then quantified based on cluster size. Furthermore, this research designs a composite weight that combines cluster weight with similarity weight, providing a comprehensive evaluation of evidence reliability from both macro-level category differences and micro-level similarity dimensions. Experimental results demonstrate that the proposed method not only resolves the negative weighting problem in existing uncertainty-based weighting methods but also effectively handles highly conflicting evidence, showing advantages in pattern recognition and other application scenarios.

One of the most common forms of audio forgery is copying and moving certain audible segments of audio to other locations in the same audio. The audio features of the pasted regions in such audio forgeries become very dissimilar to the audio features of the copied segments after post-processing. This dissimilarity makes detecting such tampering a major challenge. To address this problem, we propose a robust audio copy-move forgery detection method using a Decreasing Convolutional Kernel Neural Network (DCKNN), data augmentation, and digital fusion. In the proposed algorithm, Mel spectrogram and Hilbert–Huang spectrogram of the audio are extracted, and then they are fused by weighting coefficients, which are gained through extensive experiments. New spectrogram images are generated by weighted fusion, and these spectrogram images are used to train the proposed DCKNN model. The trained DCKNN can effectively detect copy-move forgery. The DCKNN model consists of a combination of four convolutional groups, each with different sensitivities to the two audio categories. We solve the problem of different sensitivities by sequentially lowering the parameters of the convolutional layers in the four convolutional groups, thus obtaining high accuracy in audio classification. The experimental results show that the proposed scheme is robust to most typical post-processing operations, including additive noise, compression, median filtering, resampling, re-quantization, and low-pass filtering, etc al. In addition, our method shows better performance in the detection of forged audio with multiple attacks. Compared to the state-of-the-art algorithms, the proposed algorithm has advantages in terms of accuracy, precision, and F1 score.

Digital breast tomosynthesis (DBT) combined with full-field digital mammography (FFDM), known as the “combo-mode”, can enhance breast cancer detection and discrimination. However, the DBT is not yet a standard breast cancer screening modality in most hospitals, and the “combo-mode” also doubles the dose exposure to the patient more than FFDM alone. In this study, we synthesized DBT images from FFDM to reduce the extra radiation dose and explored a methodology to effectively integrate multifaceted information from both the synthetic DBT and real FFDM. An improved conditional generative adversarial network (cGAN) network was proposed for generating synthetic DBT with image quality qualified for breast mass discrimination. A novel multiple accuracy metrics scoring (MAMS) strategy was proposed for integrating multichannel and multimodality imaging information within a hierarchical fusion framework. We retrospectively collected 441 patients with both DBT and FFDM from Nanfang Hospital (NFH) and 143 patients with only FFDM from the Second Affiliated Hospital of Guangzhou University of Chinese Medicine (GDHCM), with regions of interest (ROIs) covering the malignant, benign, and normal tissues extracted for model training and validation. The synthesized DBT exhibited satisfactory image quality and comparable discrimination ability with the real DBT on the NFH dataset. The proposed MAMS achieved an accuracy of 80%, 79%, 91%, and AUC of 87%, 85%, and 99%, respectively, for the malignant, benign, and normal tissues on the GDHCM dataset. The experimental results indicate that the findings of this study can help improve breast mass discrimination when only unimodal FFDM is available.

Salient object detection (SOD) is a key research direction in the field of computer vision and has attracted extensive attention from scholars in this field. Although deep learning has made significant progress, two key bottlenecks remain: (1) Existing methods fail to reconcile precise edge detection in low-level features with semantic coherence in high-level features, resulting in compromised boundary integrity for complex-shaped objects. (2) Conventional architectures exhibit inherent sensitivity to appearance variations due to their spatial domain limitations, and lack frequency-adaptive robustness. To address these issues, this paper proposes a novel lightweight spectral transform framework (STNet) for SOD. First, a multi-feature fusion network is introduced as a baseline model for saliency inference. An edge-guiding module is used to extract precise boundaries via differential pooling, and a semantic fusion module aligns cross-level features via dynamic dilated convolutions , both of which are integrated into this network. The objective of this integration is to efficiently aggregate fine-grained visual features and abstract semantic information while guaranteeing feature space consistency. Second, to enhance the robustness of salient information, we incorporate a spectral transform module that combines spatial and frequency domain features. This module highlights target details in multi-frequency domains and improves saliency prediction. Finally, a simple yet effective optimized loss function is designed to refine the saliency predictions. Extensive experiments confirm that the proposed STNet outperforms competing methods and is capable of accurately detecting large targets, multiple targets, and objects in both simple and complex scenarios. It achieves MAEs (Fβ) of 0.044 (0.813), 0.040 (0.914), 0.033 (0.898), 0.072 (0.821), 0.056 (0.752), 0.019 (0.975), 0.020 (0.975), 0.153 (0.842), and 0.219 (0.762) on nine benchmark datasets. Through quantitative and qualitative comparisons with state-of-the-art approaches, the proposed STNet demonstrates superior and competitive performance while retaining only 7.84M parameters and 3.73G FLOPs.

Multivariate time series (MTS) classification is a crucial research area with broad applications in action recognition, healthcare, and system monitoring. Existing methods show good performance in capturing the temporal dependence of MTS, but they fail to capture the correlation between Different sEnsors at Different Timestamps (DEDT). Meanwhile, most methods ignore the frequency-domain information of the time variables, limiting the model’s capability. To overcome the above challenges, we propose the MH-TFFN model, a novel architecture integrating selective state spaces with adaptive hypergraph learning, which achieves spatio-temporal relationship modeling of multivariate time series through time-frequency dual-channel learning mode and three breakthroughs. First, a weight sharing Mamba (WSMamba) network replaces conventional sequential processing with state-space-guided feature extraction, operating simultaneously on raw time sequences and their frequency representations obtained through Fourier decomposition. Second, an adaptive hypergraph constructor dynamically establishes DEDT relationships through sliding-window correlation analysis, subsequently processed by our time-frequency hypergraph neural network (TFHGNN), which preserves both topological and time-frequency characteristics. Third, a contrastive learning mechanism employs time-frequency adversarial pairs enforces feature consistency across domains, using a novel InfoNCE-based contrastive loss to optimize the joint space. The results of the experiment demonstrate that our proposed model outperforms state-of-the-art methods across five MTS datasets.

Backdoor attacks present significant risks to the security of deep neural networks (DNNs) in NLP domain, as the attackers can covertly manipulate the model’s output behavior either by poisoning the training data or tampering model’s training process. This paper introduces a novel backdoor defence strategy, Backdoor Defense via Ensemble Knowledge Distillation (BDEKD), to mitigate various types of backdoors in compromised DNNs. It is marked as the first utilization of ensemble methods in enhancing backdoor mitigation. The BDEKD framework only requires a minimal subset of clean data to clean the compromised model, generating several relatively heterogeneous and backdoor-cleaned teacher models. This process is followed by an enhancement of the training data through augmentation, and the implementation of an ensemble distillation technique specifically designed to mitigate the backdoor from the model. Our empirical analysis demonstrates that BDEKD effectively lowers the success rate of six sophisticated backdoor attacks to approximately 17%, while only requiring 20% of the training data. Crucially, it preserves the model’s accuracy on clean data around 85%, ensuring minimal impact on its intended functionality. Our code is available at https://github.com/quanzhuangdefujinan/BDEKD-Research/tree/BDEKD.Graphical abstract

Building extraction is essential in applications such as urban planning and monitoring urban dynamics. In high-resolution remote sensing (HRS) images, the main body of buildings and their surrounding environments exhibit strong coupling and variability, which can easily lead to the loss and underutilization of multi-scale feature information. To address this issue, we proposed a compound multi-branch fusion model, termed MCMFformer, for building segmentation in HRS images. Firstly, we designed a Mixed Multi-Branch Feature Fusion (MMFF) module, which performs multi-dimensional weighted fusion on the feature information captured by the Transformer. By employing a U-shaped attention structure, the module enhances the dimensional representation of multi-scale features from each branch and multi-level features during the aggregation phase, thereby improving the representational capacity of the building feature information. Secondly, we employed a Supervised Attention Module (SAM) to perform supervised correction on the enhanced information, thereby suppressing redundant information generated in the front-end stages. Additionally, we designed a Mixconv to mitigate information conflicts between attention branches, achieving an optimal balance point across different dimensions. We conducted comparative experiments with several state-of-the-art models on the three datasets, including Massachusetts dataset, the INRIA dataset, and the NZ32km2 dataset. On the Massachusetts dataset, the MCMFformer model achieved an average mIoU improvement of 1.44%, 2.65%, 2.13%, and 1.04% over the mainstream building segmentation models CGSANet, Segformer, MANet, and HDNet, respectively. Experiments show the effectiveness of the MCMFformer model in the task of HRS image building segmentation.

Bayesian optimization (BO) has evolved from single-agent optimization to multi-agent collaborative optimization, namely Federated Bayesian Optimization (FBO), aimed at collaboratively improving the optimization performance of all agents. Due to the limited raw data contained by each agent and privacy concerns, it is very challenging for existing FBO to perform collaborative optimization. In this work, we propose an innovative FBO method that suggests transmitting the predicted values of surrogate models between the agent and the server and aggregating the weighted model output based on the similarity of the agent to address privacy concerns. The similarity between agents is measured according to the predictions of the agents’ surrogate model. Additionally, we propose augmenting the limited raw data problem by adopting the generative adversarial network to generate a set of solutions with global information to improve the effectiveness of model management. This model output-based FBO method demonstrates competitiveness in both benchmark and real-world problems while guaranteeing privacy protection.