Applied Intelligence最新文献

With the continuous development of deep learning, generic object detection techniques are becoming increasingly mature. However, in densely populated scenes where pedestrian density is high and occlusion is severe, the performance of generic object detectors is not ideal. In dense scenes, features from the corner regions could help detectors achieve higher performance, yet existing generic detectors often overlook this aspect. To address the above issues, a plug-and-play Corner Feature Regulate Network, named as CFRN, is proposed in this paper. It extracts the corner features from the deep-layer feature maps of Feature Pyramid Network (FPN) through the Explicit Visual Center (EVC) module. Then the deep layer corner features are utilized to adjust shallow-layer features to ensure that all feature maps contain crucial corner feature essential for dense scenes. Additionally, to address the issue of redundant features which introduced by using bilinear interpolation for multi-scale upsampling in the CFRN, the Redundant Feature Suppression Module (RFSM) is proposed by using ScConv to extract redundant attention from the feature maps of CFRN. This module could reduce the redundant features introduced by using bilinear interpolation in CFRN effectively. The experiment results on the CrowdHuman, CityPersons and COCOPerson show that compared to the Sparse R-CNN, the proposed method improves by 0.7% on AP, decreases by 0.2% on (MR^{-2}), and improves by 1.1% on JI. Code is available at https://github.com/davidsmithwj/CS-CS-RCNN.

Manual testing of the mechanical parts of a satellite, such as reaction wheels, is often cost-inefficient, with multiple approaches for automating this process having been investigated. AI-driven testing tools can be employed towards this direction, reducing time and effort, although human intervention must be still kept in the loop, to ensure test robustness and coverage completability. This paper employs machine learning models, with the aim to automate the test generation process for a reaction wheel, in the context of a passive rundown test scenario. Towards this direction, a relevant simulator is utilized, specifying different models for different wheel types. A Support Vector Machine with RBF Kernel is trained guided by an efficient Active Learning technique, in order to generate adequate passive rundown test cases for three reaction wheel profiles (a heavy, a medium and a light wheel). The model is supported by an automated test evaluation script that 1 Final revised PDF analyses simulation speed outputs, labeling each test case as “adequate” or “not adequate”. The proposed approach is evaluated against a big dataset of labeled test cases for all wheel types. The experimental evaluation illustrated a high precision (1.0, 0.85 and 0.70 for the three reaction wheel profiles respectively) achieved by the SVM-RBF approach, in contrast to other competing methods (generative or classification). Lastly, as a practical experiment, four generated test cases are ran in a lab reaction wheel, estimating its friction parameters, in order to assess its health status. Calculating the observed friction values based on the time to stop parameter, the reaction wheel is finally characterized as healthy.

This research investigates the efficacy of advanced deep reinforcement learning algorithms in the field of unmanned aerial vehicle (UAV) attitude control. Deep reinforcement learning has demonstrated encouraging results in the robotics and high-level mission planning, its application to low-level mission planning remains understudied. There is a crucial need for more sophisticated control in UAV to operate in challenging and unpredictable environments. In this research study, we present a thorough assessment of the three groundbreaking deep reinforcement learning algorithms such as Proximal Policy Optimization (PPO), Deep Deterministic Policy Gradient (DDPG) & Trust Region Policy Optimization (TRPO) for UAV attitude control, quadcopters to be precise. We have utilized a high- fidelity simulation environment and experimentally assessed these algorithms’. ability to learn robust control policies. Evaluations based on rise time, peak response, error minimization and stability demonstrate that PPO outperforms DDPG and TRPO, achieving better precision and robustness. Our comparative results suggest that the controller with combination of adaptive gain optimization and DRL significantly improves the flight performance with an average of 19.9%, 6.6%, 16.5% improvement in rise time, peak percentage, error rate respectively and a complete stable flight.

Top-k high utility itemset mining(Top-k HUIM) has emerged as a critical research area, facilitating the discovery of valuable itemsets without predefined thresholds. Existing methods primarily focus on datasets without negative utilities, while approaches for handling negative utilities remain limited. Additionally, many top-k HUIM techniques require multiple global scans and large data structures, which hinder their efficiency and scalability. To address these challenges, we propose two novel algorithms: PFH (Particle Filter-based top-k HUIM for datasets without negative utilities) and PFHN (Particle Filter-based top-k HUIM for datasets with Negative utilities). PFH introduces a novel transmission process by assigning transition probabilities to particles for updating their states. A criterion for particle degeneration is proposed to terminate the transmission process, and a resampling strategy is employed to mitigate particle degeneration and improve algorithmic efficiency. In order to handle datasets with negative utilities, PFHN further introduces a utility flag filtering mechanism and employs a pruning strategy distinct from PFH to enhance efficiency. Extensive experiments demonstrate that both PFH and PFHN can efficiently and accurately mine top-k HUIs, providing a novel perspective for solving the top-k HUIM problem.

Multi-attribute group decision-making (MAGDM) constitutes a pivotal methodology for resolving complex problems requiring collective expertise. However, critical limitations persist in current frameworks, including overreliance on unilateral expert weight strategies that neglect the interplay between informational objectivity and social influence, treatment of evaluation systems as structural “black boxes”, unrealistic assumptions regarding perfect expert rationality, and inequitable modification of opinions during consensus building. To address these issues, this study develops an integrated MAGDM framework that incorporates three synergistic innovations. First, we introduce a hybrid weighting mechanism that combines information entropy and quantum Bayesian networks (QBNs) to quantify the interference effects of trust propagation, and solve the combination ratio through Monte Carlo simulation. Second, a fair consensus adjustment model is constructed to optimize the distribution of opinion revisions and balance consensus reaching with the retention of original opinions. Third, to characterize experts’ risk-avoidance behavior and evaluate the internal structure and behavioral characteristics of the system, we design a parallel data envelopment analysis (DEA) cross-efficiency model and a regret-based perceived utility value (PUV). Finally, an illustrative example is presented to validate the effectiveness of the proposed approach, while its robustness and superiority are demonstrated through sensitivity analysis and comparative experiments.

The low-rank and sparse matrix decomposition problem is a hot and challenging problem in computer science. In this paper, we consider it as a nonconvex relaxation optimization problem by using a family of nonconvex functions to approximate the rank function and the (ell _{0})-norm in low-rank and sparse matrix decomposition problem, namely, generalized low-rank and sparse matrix decomposition problem. The essence of this paper is to develop an adaptive algorithm framework with parameters updating for the nonconvex relaxation problem. Firstly, we prove the equivalence between the generalized low-rank and sparse matrix decomposition problem and the regularization generalized low-rank and sparse matrix decomposition problem. This means that the optimal solution of generalized low-rank and sparse matrix decomposition problem can be exactly obtained by solving its regularization minimization problem. Secondly, we present a tractable nonconvex algorithm framework to solve the regularization generalized low-rank and sparse matrix decomposition problem. The convergence analysis of the algorithm framework is provided. More importantly, we also define a very powerful parameter-setting strategy to adapt the optimal parameters in iteration of the proposed algorithm framework. Finally, we test the proposed algorithms on some random low-rank and sparse matrix decomposition problems, and the numerical results verified the effectiveness of the proposed algorithms. In addition, we also extend the proposed algorithms to the image denoising and background modeling from surveillance video.

This study focuses on the complexity and challenges of risk assessment in Supply Chain Finance (SCF) from an Environmental, Social and Governance (ESG) perspective. Through in-depth analysis of interpretable machine learning models, their effectiveness in identifying ESG-related risk was evaluated. Firstly, a SCF risk assessment framework that integrated ESG factors was designed. Subsequently, the data were preprocessed using the synthetic minority oversampling technique (SMOTE). Then, a variety of machine learning models for risk assessment were applied. Finally, through ablation experiment and Shapley Additive interpretation (SHAP) of XGBoost (Extreme Gradient Boosting), this study explained the contribution and importance of each risk factor to the results. This study effectively solved the interpretability problem of black-box machine learning model. The results of ablation experiments showed that ESG factors had certain influence on the risk of SCF. In addition, SHAP method further highlighted the core role of asset-liability ratio, cash ratio and quick ratio in risk assessment. This study revealed the practical application effects of different models in financial risk assessment, and used SHAP algorithm to provide clearer verification of machine learning results. It filled the gap in interpretable machine learning in SCF field. And it also provided strong support for promoting green and sustainable supply chain financial resources.

This paper presents a Multimodal Dynamic Inference Framework (MDIF) for video question answering in traffic scenarios. MDIF is different from existing methods that rely on correlation modeling and global features. It combines causal inference with object segmentation. Causal inference reduces spurious correlations and ensures unbiased reasoning. Object segmentation extracts fine-grained information such as traffic signs, vehicle movements, and pedestrian interactions. Together, they improve causal modeling, scene understanding, and event prediction accuracy. MDIF also employs Graph Convolutional Networks (GCNs) to capture spatiotemporal dependencies in complex and dynamic traffic events. We evaluate MDIF on two large-scale datasets, SUTD-TrafficQA and MSVD-QA. The results demonstrate significant progress in prediction and counterfactual tasks, validating the robustness and generalization ability of the MDIF framework. Ablation studies further confirm the role of spatiotemporal modeling, object segmentation, and causal intervention. These components are all essential to robust, interpretable, and unbiased reasoning. Overall, MDIF provides strong advantages in causal modeling and cross-modal inference.

Visual Question Answering (VQA) models often answer questions based on the superficial correlations between question-answer pairs rather than actual reasoning. It leads to good performance on in-distributed (ID) datasets and a significant decline on out-of-distributed (OOD) datasets. Existing debiasing methods primarily focus on single-modal branches, with few studies addressing multi-modal bias simultaneously. A bias extraction and penalty (BEP) method was proposed in this paper. Using a generative adversarial network and knowledge distillation strategy, the bias is extracted directly from the VQA model and incorporated into the bias model. Furthermore, margin penalty is introduced to represent the frequency of certain answer types and the difficulty of sample answers as margin information. The size of the margin reflects the degree of bias, and different penalties are assigned to samples with varying degrees of bias. Supervised contrastive learning is employed to retain these penalties, enabling the model to focus more on training biased samples. Additionally, a classifier based on Cross-Entropy(CE) loss was proposed, which has a stronger inference ability on ID datasets, and uses the main classifier and CE loss classifier for joint inference. Experiments on the challenging VQA-CPv2 and VQA v2 datasets show that BEP achieves state-of-the-art results among non-augmentation debiasing methods while maintaining competitive performance on ID datasets.

In video anomaly detection tasks, multi-modal input can encode video content across different feature spaces, offering richer and more complementary semantic representations than single-modal data. However, most existing multi-modal VAD methods employ independent modeling strategies and fuse features at the decision stage typically through simple operations such as concatenation or gating. To address this limitation, we propose a novel Multimodal Interaction and Fusion Mechanism (MMI-FM) based on RGB features and skeleton features. Specifically, we design the Dual Contrastive Loss Alignment Module (DCL-AM), which aligns features by mining both intra-modal and cross-modal semantic correlations. Furthermore, we introduce a Cross-Modal Bidirectional Knowledge Distillation Module (CM-BKDM) to mitigate potential representational biases in single modality representation learning by performing bidirectional knowledge transfer between modalities. Finally, an Adaptive Multimodal Feature Fusion Module (AMFFM) is proposed, which dynamically fuses RGB regions and skeleton points with strong semantic associations. Extensive experiments conducted on two public VAD datasets, CUHK Avenue and ShanghaiTech, demonstrate that MMI-FM achieves AUC scores of 91.8% and 78.4%, outperforming most state-of-the-art methods and validating the effectiveness of our proposed framework.