Applied Intelligence最新文献

Dataset ownership verification (DOV) enables an individual to confirm their ownership of the training dataset for an AI model. DOV is particularly valuable in situations where a model is believed to have used copyrighted datasets for training without obtaining permission. Invisible backdoor watermarking provides proof of ownership through human-invisible backdoors. However, real-world AI environments like MLaaS may deploy defense models that deactivate both adversarial backdoors and useful watermarks. In this paper, we propose a noble DOV method that improves the stealthiness of invisible backdoors. Specifically, invisible backdoors with high attack success rate (ASR) are typically detected, whereas those with low ASR are undetectable. In our scheme, we prove how to employ backdoors with low ASR but at the same time achieve significantly higher DOV success rates. This is accomplished through the training of two distinct models: a watermarked model and a verification model. The verification model is trained and tested using the differences between the output confidence vectors, where these vectors are obtained by inputting watermarked and clean image pairs into watermarked models. With this approach, we achieve on average a high DOV success rate, 94.61% in four representative image datasets, the CIFAR10, CIFAR100, GTSRB, and Tiny ImageNet dataset.

Accurate segmentation of cardiac structures remains a challenging task in cardiac image analysis, particularly due to complex anatomical variations and dynamic, nonlinear deformations across the cardiac cycle. Traditional approaches often struggle to maintain precision under these conditions, especially when faced with ambiguous boundaries and low-contrast regions in cardiac magnetic resonance imaging (MRI). To handle this, we explore the hierarchical nature of cardiac deformations while preserving structural boundaries and propose DABC-Net (Deformation Aggregation Boundary-Constrained Network), a novel framework that integrates hierarchical deformation feature aggregation with boundary-aware supervision to enhance cardiac MRI segmentation. To effectively encode the spatially varying deformation patterns, we design SDIM (Shape Deformation Integration Module) that progressively models multi-level anatomical deformations, enabling the network to adaptively learn deformation-aware representations across different feature stages. Complementarily, a DBAM (Dual-Branch Adaptive Mixer) is introduced to bridge fine-grained feature and broad contextual semantics, promoting robust alignment in the presence of complex shape dynamics and intensity inconsistencies. To further address the challenge of precise boundary localization, we incorporate a Boundary-Constrained Supervision Strategy, which guides the network to focus on fine structural details through both architectural and loss-based refinements. Extensive experiments on publicly available MM-WHS and ACDC datasets demonstrate that DABC-Net consistently achieves high segmentation accuracy and robust generalization across different cardiac pathologies. Our results underscore the effectiveness of combining hierarchical deformation modeling with structural boundary constraints, achieving state-of-the-art performance while advancing anatomical understanding in cardiac image segmentation.

Chinese ancient paintings have suffered from human damage or accidental fires, resulting in large-area damage to the paintings. However, the natural image restoration technology is not applicable to ancient paintings. Therefore, a large-area damage inpainting model is proposed for ancient paintings based on Spatial Fourier Convolution. Firstly, Spatial Fourier Convolution is constructed for large-area damage inpainting. In Spatial Fourier Convolution, Spatial Mixed Attention (SMA) is proposed to alleviate artifacts and unreasonable color filling. In addition, Multi-Scale Dilated Attention (MSDA) is introduced in spatial Fourier convolution to effectively capture the multi-scale information. Secondly, the Region Normalization (RN) is introduced to eliminate the artifacts produced due to the loss of feature information in the broken area. Then, in order to address the problem of the loss of line structure information, edge loss is incorporated to retain the original structure information, so as to accurately reconstruct the contours and details of the ancient paintings. Lastly, a large-area damage dataset is created for ancient paintings, and experiments are conducted to evaluate the performance of the proposed model. Qualitative and quantitative experiments show that, compared with the latest inpainting models, the proposed model significantly reduces loss of details, blurred textures, missing lines, and unnatural color reconstruction. Therefore, it is capable of inpainting the large-area damage of ancient paintings.

Graph neural networks (GNNs) are widely utilized for modeling graph data. However, most existing GNNs are trained in a task-driven manner, focusing on proximity learning, which fails to fully capture the intrinsic nature of the graph structure and results in suboptimal node and graph representations. To address this issue, we present a novel and generalized training method called graph structure prompt learning (GPL), inspired by prompt mechanisms in natural language processing, to enhance GNN training. GPL employs task-independent graph structure losses to encourage GNNs to learn intrinsic graph characteristics while simultaneously solving downstream tasks, thus producing higher-quality node and graph representations. Extensive experiments conducted on eleven real-world datasets and 20 GNNs demonstrate that GNNs trained with GPL significantly outperform their baselines in node classification, graph classification, and edge prediction tasks, achieving improvements of up to 10.28%, 16.5%, and 24.15%, respectively. By enabling GNN to capture the inherent structural prompts of graphs in GPL, GPL also mitigates the over-smooth issue and achieves new state-of-the-art performance benchmarks, paving the way for innovative research directions in GNNs with potential applications across various domains.

With a user-specified minimum utility threshold (minutil), periodic high-utility pattern mining (PHUPM) aims to identify high-utility patterns that occur periodically in a transaction database. A pattern is deemed periodic if its period aligns with the periodicity constraint set by the user. However, users may not be interested in all periodic high-utility patterns (PHUPs). Moreover, setting minutil in advance is also a challenging issue. To address these issues, our research introduces an algorithm called TPU for extracting the most significant top-k periodic and high-utility patterns that may or may not include negative utility values. This TPU algorithm utilizes positive and negative utility lists (PNUL) and period-estimated utility co-occurrence structure (PEUCS) to store pertinent itemset information. Additionally, it incorporates the periodic real item utility (PIU), periodic co-occurrence utility descending (PCUD), and periodic real utility (PRU) threshold-raising strategies to elevate the thresholds rapidly. By using the proposed threshold-raising strategies, the runtime was reduced by approximately 5% on the datasets used in the experiments. Specifically, the runtime was reduced by up to 50% on the mushroom_negative and kosarak_negative datasets, and by up to 10% on the chess_negative dataset. Memory consumption was reduced by about 2%, with the largest reduction of about 30% observed on the mushroom_negative dataset. Through extensive experiments, we have demonstrated that our algorithm can accurately and effectively extract the top-k periodic high-utility patterns. This paper successfully addresses the top-k mining issue and contributes to data science. Furthermore, the applications of the proposed algorithm in engineering include data mining, expert systems, and web intelligence in various fields, such as smart retail, cyberspace security, and risk prediction. The code and datasets are publicly available at https://github.com/DSI-Lab1/TPU.

As an important image preprocessing method, salient object detection has achieved excellent performance in various computer vision tasks. Moreover, salient object detection under weak observation conditions has deeper research value and application prospects in the fields of assisted driving and autonomous driving. However, existing salient object detection methods are generally designed for high-quality images obtained in conventional good scenes. It is a great challenge for these methods to solve the problem of difficulty in excavating the saliency information of foreground objects in low-quality images caused by occlusion and interference from specific weather background. Therefore, a novel unsupervised adaptive learning method for salient object detection under weak observation conditions is proposed, which generates accurate saliency maps by constructing a pseudo label generation network and a saliency detection network. Specifically, the adaptive learning of filtering parameters module is designed in the pseudo label generation network to augment image quality and assist in excavating saliency information by dynamically remove interference from adverse backgrounds. In addition, the saliency feature grafting module that can fuse low-level and high-level semantic features of images is proposed in the saliency detection network to refine object edges. The experimental results show that the proposed method adaptively detects salient objects under normal and weak observation conditions.

To suppress the extreme fluctuations of pantograph-catenary contact force (PCCF) taking account of wheel-rail excitation of high-speed trains, a real-time active control framework based on a hybrid deep reinforcement learning (DRL) algorithm is proposed. Specifically, the influence of wheel-rail excitation on the vertical vibration amplitudes of the pantograph is clarified by establishing the multi-body dynamic model based on the interaction mechanism between the rail-vehicle (RV) system and the pantograph-catenary (PC) system. Subsequentially, a hybrid framework based on the feedback control (FC)-guided DRL algorithm is established for applying active control force to the pantograph. A full-state Linear Quadratic Regulator (LQR) is implemented in the FC section to optimize the primary dynamic states of the train, with the forces applied to the pantograph play the guiding role. For more nuanced policy exploration, a self-attention-enhanced DRL-based controller employs the Soft Actor-Critic (SAC) algorithm with the Hindsight Experience Replay (HER) mechanism to train a policy aimed at eliminating extreme PCCF values. The reward function designed for time-varying speed conditions, along with the structure of the FC-DRL, is detailed extensively. Compared with the baseline algorithm, the simulation results based on random track irregularity input show that our control scheme reduces the standard deviation of PCCF by a maximum of 41.85% and the average error by a maximum of 78.25%. The proposed framework converges stably, and the extreme contact force can be eliminated effectively in various testing environments.

Early wildfire detection is important for managing fire to avoid severe damage to ecosystems and humans. Recently, computer vision-based fire detection approaches have become of interest to various researchers as they allow the monitoring of large areas, facilitating fire early detection. Generally, fire color is the main characteristic used in various rule-based approaches. Rule-based approaches determine fire pixels by assessing whether the rules are satisfied by exceeding a threshold value determined previously. In this research, we propose a rule-based wildfire pixel detection approach without thresholds. Our method estimates contrast characteristics of fire in various color spaces and combines the extracted features in a simple but still effective rule-based manner. The proposed approach consists of two main stages. Firstly, from various color spaces, the color-contrast estimation is performed and encoded into a set of gray-level images. Secondly, our proposed method combines various contrast features to obtain a final fire image. In contrast to the other methods, where a threshold is specified for the determination of fire pixels, our approach determines the fire pixel by using merely the color contrast features computed from the incoming image. We evaluated our proposal with ten rule-based and two learning-based fire detection methods. The results indicate that our proposal achieves competitive performance in three evaluation metrics.

Explaining machine learning (ML) models using eXplainable AI (XAI) techniques has become essential to make them more transparent and trustworthy. This is especially important in high-risk environments like healthcare, where understanding model decisions is critical to ensure ethical, sound, and trustworthy outcome predictions. However, users are often confused about which explanability method to choose for their specific use case. We present a comparative analysis of two explainability methods, Shapley Additive Explanations (SHAP) and Gradient-weighted Class Activation Mapping (Grad-CAM), within the domain of human activity recognition (HAR) utilizing graph convolutional networks (GCNs). By evaluating these methods on skeleton-based input representation from two real-world datasets, including a healthcare-critical cerebral palsy (CP) case, this study provides vital insights into both approaches’ strengths, limitations, and differences, offering a roadmap for selecting the most appropriate explanation method based on specific models and applications. We qualitatively and quantitatively compare the two methods, focusing on feature importance ranking and model sensitivity through perturbation experiments. While SHAP provides detailed input feature attribution, Grad-CAM delivers faster, spatially oriented explanations, making both methods complementary depending on the application’s requirements. Given the importance of XAI in enhancing trust and transparency in ML models, particularly in sensitive environments like healthcare, our research demonstrates how SHAP and Grad-CAM could complement each other to provide model explanations.

To address premature convergence, slow convergence, and parameter sensitivity in conventional ant colony algorithms for static optimization, we propose an elite quantum ant colony algorithm based on double chain encoding (DE-QACA). The algorithm employs a sine/cosine, real-valued pheromone representation that explicitly decouples exploration from exploitation. Adaptive quantum rotation angles, triggered by objective improvement, guide search, while a quadratic-decay elite pool mitigates late-stage stagnation and improves robustness to parameter settings. We establish convergence guarantees and derive time/space complexity bounds. Evaluations on Traveling Salesman Problem (TSP) instances and CEC2017 continuous benchmarks show that DE-QACA attains higher success rates on large-scale TSP and converges faster on hybrid functions than competitive baselines, demonstrating fast convergence across discrete and continuous domains.