Applied Intelligence最新文献

With the rapid advancement of deep learning and computer vision, few-shot object detection (FSOD) has emerged as a critical research frontier. A key challenge in FSOD lies in extracting discriminative feature representations from limited samples, which severely degrades detection performance. To mitigate this issue, we propose a novel FSOD framework that integrates cross-domain adaptive feature enhancement and attention-guided proposal generation, effectively leveraging support set information to improve query set detection accuracy. Our method introduces three key innovations. (1) Dynamic Kernel Generation. A learnable kernel generator produces sample-specific convolutional kernels to adaptively enhance query features using support set cues. (2) Attention-Driven Region Proposals. An attention-based region proposal network (ARPN) suppresses irrelevant regions while prioritizing semantically relevant areas. (3) Template-Aware Scoring. A matching module evaluates candidate boxes against support templates to ensure geometric and semantic consistency. Extensive experiments on PASCAL VOC and MS COCO benchmarks demonstrate our method outperforming existing approaches by 3.2 AP50 on 10-shot tasks. The results validate the efficacy of cross-domain adaptation and attention mechanisms in addressing data scarcity challenges.

Until now, a number of one-class classification methods have been used for anomaly detection problems. However, only single one-class classifiers often fail to accurately detect anomalous patterns because they might not be optimized to describe inherent target class structures. To achieve more superior anomaly detection performance, we propose random neighborhood ensemble-based one-class classification algorithm. In the proposed algorithm, multiple one-class classifiers are obtained from various random k-NN sets identified by ensemble of random subspace and random neighborhood size approaches, and they are finally aggregated as novelty score. The random subspace and random neighborhood size approaches helps to diversify k-NN sets, and these diversified k-NN sets can effectively accommodate the inherent target class structures having complex patterns. In this study, we conducted experimental studies with various benchmark datasets to investigate the characteristics of the proposed algorithm and compare it with existing methods. The experimental results demonstrate that the proposed algorithm performs better or comparable to existing one-class classification methods in most of cases.

In three-way conflict analysis, alliance, neutrality, and conflict reducts derived from three-valued situation tables play a crucial role in identifying and resolving potential conflicts. However, existing research has largely focused on standard three-valued situation tables, with insufficient attention paid to three-valued decision situation tables. To address this gap, we first introduce conflict measures regarding both conditional and decision issues, and establish alliance, neutrality, and conflict relations with respect to these two types of issues within a three-valued decision situation table. We then provide the concepts of relative alliance, neutrality, and conflict reducts tailored to three-valued decision situation tables. To facilitate the systematic construction of these reducts, we formulate three types of relative discernibility matrices and design an algorithm for computing the corresponding reducts based on these matrices. Finally, the practical applicability of the proposed approach is demonstrated through two case studies: the NBA labor negotiations and the development planning of Gansu Province. A comparative analysis highlights the effectiveness and advantages of the proposed method compared to existing approaches.

In group decision-making problems related to sudden events, the absence of evaluation attributes and the emergence of minority opinions are common problems. Due to the shortcomings of existing methods for such problems, this paper proposes a group decision-making model based on public big data and minority opinions for sudden events. Firstly, a linear algorithm for the comprehensive attribute weights is novelly constructed based on two aspects of the public level and the expert level, the attribute weights of the public level are derived from a normalization algorithm by combining the term frequency-inverse document frequency technology with the latent dirichlet allocation model, and the attribute weights of the expert level are determined by constructing a maximum nonlinear deviation model. Secondly, by incorporating both the support degree and trust degree in the adjustment coefficient, a new valuable minority opinion management scheme is proposed to make the decision-making more reasonable and objective. Thirdly, a consensus optimization model is established based on minimum adjustment to protect minority opinions and improve decision-making efficiency. Finally, an application is demonstrated to illustrate the efficiency and flexibility of the proposed model.

Learning representations of users and items are importance of predicting user preferences for accurate recommendation. However, confusion features of users genuine interests and conformity behaviors often hamper the accurate representation of user preferences and lead to inappropriate recommendation. The existing methods overlook the fine-grained attributes of interest and conformity, hiding the actual preferences of users. In this paper, we propose a similarity popularity-based adaptive disentangled learning recommendation model with individual preference to learn representations from the observed data with the confusion features of interest and conformity. Firstly, a novel popularity calculation method based on item similarity is presented, which better captures fine-grained attributes and distinguishes user interest and conformity to generate more accurate sampling signals. Subsequently, we design a weight attention score mechanism that dynamically adjusts the weights of interest and conformity, accurately aligning with the individualized preferences of users. Finally, we develop a comprehensive loss function to calibrate the importance of disentangled representation learning. Experiments on public datasets demonstrate that the proposed model outperforms the existing debiasing methods, mitigating the popularity bias of recommendations from fine-grained attributes perspective.

To address equipment aging, spatial signal redundancy, and robustness deficiencies in long-term operation, this paper proposes a cylindrical-coordinate Quantized Iterative Learning Control framework (QILC-CC). The method exploits geometric symmetry to quantize multi-dimensional signals and reduce communication load, introduces a dynamic learning gain to accelerate convergence, and embeds an aging-compensation term to suppress slow degradation, thereby unifying accuracy and stability within a single learning law. Theoretically, we derive the error recurrence under bounded disturbances and gradual aging, establish sufficient convergence conditions and steady-state error bounds, and justify them via Lipschitz continuity and Lyapunov energy arguments. In simulations on a tower-type solar-thermal heliostat field, QILC-CC improves convergence speed by approximately 50-60%, reduces steady-state error by 20 – 45%, and, under strong disturbances, lowers RMS error by 25% ; under equal-accuracy alignment, it further reduces communication and computation burdens by 30 – 40%. Multiple independent runs with statistical validation (R = 30, p ˂ 0.01) confirm the significance and robustness of these gains. Comparative studies against MRAC, L₁ adaptive control, H∞ control with ESO, and related advanced methods show that QILC-CC delivers superior long-term steady-state performance and resource efficiency for repetitive tasks with slow aging. Taken together, these results position QILC-CC as a low-bandwidth, low-compute, and statistically verifiable control solution for long-horizon operation.

Accurate photovoltaic (PV) power forecasting is essential for power system scheduling and security, yet it is challenged by the high uncertainty of solar generation due to weather, season, and location. Although Transformer-based models have advanced prediction capabilities, they still suffer from key limitations: their point-wise self-attention underuses local context, they capture short-term dependencies poorly, and their numerous hyperparameters require experience-heavy manual tuning, complicating optimization. To overcome these issues, this paper proposes a novel Transformer-based model named MSLP-Gruformer for PV power forecasting. Built upon the iTransformer framework, the model incorporates a Multi-scale Local Perceptual Attention Mechanism (MSLP) to enhance the extraction of local features and trend variations in time-series data. Furthermore, by integrating a Gated Recurrent Unit (GRU) module, the model strengthens its ability to jointly capture both long- and short-term dependencies. Hyperparameter optimization is automated using the Blood-Sucking Leech Optimizer (BSLO), which reduces computational cost and improves tuning efficiency. The performance of MSLP-Gruformer is evaluated using real-world data from four PV stations located in China and Australia. Forecasts are generated for four time horizons: 1 h, 4 h, 12 h, and 24 h, and compared with the baseline iTransformer model. Compared with GRU, TCN, Transformer-GRU, Transformer-TCN, Crossformer and iTransformer, results show that MSLP-Gruformer achieves consistently lower prediction errors and higher goodness-of-fit across all forecasting horizons and station capacities, demonstrating its potential to support daily power system scheduling and enhance operational security.

Cultural relic image restoration presents unique challenges due to irregular damage and historically specific textures, which standard deep learning methods struggle to address. This paper proposes a novel two-stage Transformer-CNN framework tailored for this task. The first stage leverages a Transformer to capture global structural dependencies from low-resolution priors, generating coherent coarse proposals. The second stage employs a specialized CNN to refine fine-grained textures from these proposals, optimized by a compound perceptual loss function. Validated on a new large-scale dataset of 88,000 East Asian cultural relic images, our approach demonstrates state-of-the-art performance. A key contribution is the generation of diversified restoration outputs, providing conservators with multiple valid references for decision-making. This work establishes an effective paradigm for digital heritage conservation that balances global structural integrity with local texture fidelity.