Applied Intelligence最新文献

Video anomaly detection aims to detect incidental and sudden patterns of events in complex scenes, which plays an essential role for maintaining public security and averting potential risks. Unsupervised learning is particularly effective in solving the problem of insufficient annotation of video abnormal behaviour in low resource scenarios. However, due to the complex entangled factors contained in the appearance/motion features of videos, it remains a formidable task for unsupervised video anomaly detection to overcome the barrier of spatio-temporal coupling. Most related methods focus on constructing a compact normal manifold within a normality training dataset or tracking objects, which often neglect the distinct visual characteristics of various anomalies and lead to the static coupling bias. To unravel the spatio-temporal coupling mechanism, we present an unsupervised approach STVAD that focuses on efficient context modelling and elimination of the spatio-temporal factor gap in a novel encoder-decoder network. Additionally, we propose a new paradigm that integrates spatio-temporal dependencies using self-attention mechanisms to boost the discriminative capacity to identify human-related irregular events from surveillance video sequences. Extensive experiments validate the accuracy and computational effectiveness of our method on three challenging benchmarks: UCSD Pedestrian(Ped1 and Ped2), CUHK Avenue and ShanghaiTech datasets respectively. Notably, our method significantly exploits the spatio-temporal clues to estimate association and improves the detection accuracy in detecting anomaly behaviours.

The application of new coding standards brings a wider range of applications as well as new video forensic challenges. In order to solve the transcoding detection problem under the new video standard H.266/Versatile Video Coding (H.266/VVC), we propose a new algorithm based on Coding Unit (CU) block type classification and Motion Vector Map (MVM) statistics to detect VVC transcoded videos. We first analyze the generation of HEVC/VVC transcoded video traces and define MVM graphs for constructing features. By calculating the average frequency of different CU partition types for intra-frame prediction frames and inter-frame prediction frames, they are connected and sent as distinguishing features to a support vector machine for classification. Experimental results show that the method has high recognition accuracy under the new coding standard. Specifically, the algorithm constructs a 51-dimensional feature vector by combining intra-frame and inter-frame features, capturing transcoding traces such as changes in CU partitioning and motion vector distributions. Extensive experiments on 1080p and 4K video datasets, across various bitrate combinations, quantization parameters, and Group of Pictures (GOP) structures, demonstrate an average detection accuracy of 96.81% on 1080p datasets, outperforming existing methods. The robustness of the proposed approach across diverse encoding configurations highlights its effectiveness for forensic analysis under the VVC standard.

The ability to predict blunders in chess plays a crucial role in improving players’ performance and enabling strategic decision-making. We introduce a novel, scalable, and personalized blunder prediction model for chess. Unlike prior work requiring a separate model per player, our unified architecture learns a collaborative user embedding space, allowing it to generalize weaknesses across players and new users. Our hybrid model, inspired by Deep Factorization Machines (DeepFM), fuses a frozen pre-trained CNN (for board embeddings) with dynamically learned user embeddings to model player-board interactions while still utilizing metadata about the state of the board and the user. We demonstrate that this latent ’blunder profile’ is a significantly more powerful predictor of error than a player’s explicit Elo rating. The system achieves state-of-the-art performance (0.801 AUC) on both immediate and non-immediate blunders, offering an efficient and data-sparse-friendly solution for personalized chess analysis. Ultimately, this approach demonstrates the practical viability of deep personalization in complex strategy games, facilitating highly efficient, user-centric learning environments.

Existing approaches for multimodal hand feature fusion typically involve extracting features from different modalities independently and subsequently merging them directly. However, these methods overlook potential correlations between modalities, leading to limitations in effectively leveraging complementary information across modalities. Therefore, this paper proposes a multimodal hand feature fusion method based on modality-related feature interaction (MR-FIFM). Firstly, a specific-shared feature extraction network(SSFEN) is constructed, preserving the integrity of features of each modality while enhancing the correlations between modalities. Secondly, a class-constrained multimodal metric loss is proposed to avoid the gradient vanishing problem and enhance the category representation capability of features. Finally, a feature interaction fusion module(FIFM) is proposed, allowing the model to simultaneously focus on features at different positions within each modality to enhance intra-modality features. Simultaneously, the features between the two modalities are interacted, enabling the model to dynamically adjust its own feature representation based on the features of the other modality. Experimental results on public datasets, including palmprint + palm-vein (PolyU_Palmprint+PolyU_NIR, PolyU_Blue+PolyU_NIR), and fingerprint + finger-vein (NUPT_FP+NUPT_FV), demonstrate that the proposed method outperforms existing multimodal fusion recognition methods for hand features.

Online object detection is a crucial task related to the safety of autonomous driving during dynamic movement. While previous researches have achieved a significant breakthrough in low-latency detection by pursuing higher FPS, there has not been a clear analysis between detection speed and online object detection performance, particularly for non-real-time detectors (e.g. DETR-based detectors). To address this issue, we propose a metric named fAP to select a future frame as the matching target for the current frame according to the FPS of the detector. Through this new metric, we reduce the bias of detection performance evaluation caused by input frame inconsistency and verify the potential of the object detection model based on transformer architecture for the online object detection task. To this end, we develop a novel end-to-end object detector in conjunction with streaming perception, named Stream-DINO, which has better online object detection capability than other transformer-based models. Specifically, a Feature Encoder Acceleration Module is proposed to reduce the computation cost, thus make it focus on small target objects in each frame. We also introduce an Associated Frame Loss to utilize the object information captured in the associated frame to supervise the detection results and obtain accurate location of target objects. Notably, Stream-DINO is the first transformer-based online object detection method, and we achieve SOTA performance ( 19.9% in fAP and 18.6% in sAP) on Argoverse-HD dataset, outperforming the baseline (DINO) by 2.2% in fAP and 3.0% in sAP respectively.

Time series forecasting is a key tool in financial markets, helping to predict asset prices and guide investment decisions. In highly volatile markets like Bitcoin (BTC) and Ethereum (ETH), forecasting is challenged by sharp price swings driven by sentiment, technological shifts, and regulatory changes. Traditionally, forecasting relied on statistical methods, but as markets became more complex, deep learning models like LSTM, BiLSTM, and the newer FinBERT-LSTM emerged to capture intricate patterns. Building upon these advancements and addressing the volatility inherent in cryptocurrency markets, we propose a novel hybrid model that combines Bidirectional Long Short-Term Memory (BiLSTM) networks with FinBERT. In volatile markets like cryptocurrencies, where price movements reflect both immediate reactions and delayed responses to sentiment, the FinBERT-BiLSTM model offers a clear advantage over unidirectional approaches. We evaluate the model against traditional baselines (LSTM, BiLSTM), sentiment-aware variants (FinBERT-LSTM, FinBERT-TCN), and state-of-the-art models, including Transformer-based (Informer, TCN, TFT) and GPT-based (TimeGPT) architectures. These evaluations span short-term (intra-day and one-day-ahead) and long-term (30-day-ahead) forecasting horizons, augmented with a realistic trading simulation. Experimental results show that FinBERT-BiLSTM achieves low Mean Absolute Percentage Error values across all horizons—2.03% (BTC) and 2.52% (ETH) for intra-day, 2.20% (BTC) and 2.77% (ETH) for one-day-ahead, and 1.62% (BTC) and 5.10% (ETH) for 30-day-ahead prediction. Moreover, it outperforms all competing models in simulated trading profitability, demonstrating both statistical and practical value for sentiment-informed cryptocurrency forecasting.

This study investigates how visual design elements—specifically color, lighting, and spatial composition—affect emotional and empathetic engagement in Virtual Reality (VR) environments. It focuses on adult users across diverse domains such as education, psychotherapy, and humanitarian advocacy. A systematic review of literature from 2018 to 2023 was conducted using Scopus, Web of Science, and JSTOR, adhering to PRISMA guidelines. A total of 21 peer-reviewed studies were selected based on their focus on visual modalities in VR and empathy-related outcomes for adult users. The review identifies a consistent impact of visual elements on users’ empathetic and emotional responses in immersive VR contexts. However, it also uncovers key challenges: (1) a lack of standardized, validated metrics for measuring empathy in VR; (2) a fragmented approach to visual emotion analysis that isolates design elements rather than considering their interplay; and (3) limited investigation into cultural variability in interpreting visual stimuli. This study offers a novel contribution by synthesizing visual design practices in VR to foster empathy, a focus not previously addressed in our research. Unlike existing studies, including any prior work by our team, this review integrates cross-cultural perspectives, proposes standardized empathy metrics, and examines the interplay of visual elements (color, lighting, spatial composition) across diverse domains, providing a cohesive framework for empathy-driven VR design.

Recognizing Textual Entailment (RTE) is a fundamental task in natural language processing with broad downstream applications. In low-resource scenarios, existing methods struggle due to their reliance on external knowledge, sensitivity to semantic distribution bias, and limitations in capturing fine-grained entity distinctions. To address these challenges, we propose a novel framework that enhances semantic awareness of sequence entities through two key strategies: Sample-based Entity Replacement Augmentation and Integration of Entity Features. First, we extract and categorize entities from the dataset, then augment the data by replacing entities with others of the same type, leveraging semi-supervised learning to enhance model training. Second, we introduce the Advanced Contrastive Learning Encoder (ACE) model, which integrates filtered entity features into sequence representations and employs sequence-entity contrastive learning to improve entity distinction. Our approach significantly reduces dependence on external knowledge, mitigates semantic distribution bias, and enhances entity-awareness in RTE. Experimental results demonstrate that our methods consistently achieve state-of-the-art performance across all benchmark datasets for low-resource Chinese RTE tasks.

We propose a comprehensive system architecture designed to effectively mitigate the growing risks associated with modern DeepFake technologies. Our proposed system comprises of three primary components, each playing a vital role in the detection process. In the first component, we proposed a GAN framework with a dual discriminator, specifically developed for initial image authenticity verification. This innovative approach is essential for differentiating genuine images from those artificially manipulated. Following this, we employ a specialized form of CNN, known as a Residual Block-Based CNN, to further analyze images verified by the GANs. This deep network, consisting of 33 layers, excels in extracting and processing intricate features from the images, a critical factor in identifying complex patterns indicative of anomalies. Lastly, the system incorporates a Multiple Instance Learning (MIL) framework. The framework presented in the research significantly enhances image anomaly detection at the event level, rather than just the instant level. The system was rigorously evaluated on two benchmark datasets: FaceForensics++ (FF++) and CelebDF. It achieved a classification accuracy of 96.01% with a False Negative Rate (FNR) of 18.86% and an average computation time of 1801.10 seconds on the FF++ dataset. On the CelebDF dataset, it attained an accuracy of 94.3%, an FNR of 5.9%, and an average computation time of 110.04 seconds. These results demonstrate the proposed framework’s effectiveness and efficiency in detecting complex image forgery cases without requiring extensive manual annotations, making it suitable for real-time surveillance and media integrity verification systems.

Session-based recommendation (SBR) faces critical limitations in existing approaches that hinder recommendation accuracy. Current methods suffer from three fundamental deficiencies: insufficient modeling of temporal dependencies, where standard attention mechanisms treat all sequence positions equally and fail to capture the relative importance of temporal proximity; noisy aggregation in graph neural networks (GNNs), where multi-layer GNN architectures cause over-smoothing and feature information loss as node representations converge to indistinguishable states; and computational inefficiency, where quadratic-complexity self-attention mechanisms introduce noise through indiscriminate all-to-all item connections. To address these challenges, we propose the Mamba-Integrated Spatio-Temporal Attention Graph Convolutional Network (MSTA-GNN). MSTA-GNN tackles temporal dependency neglect through a convolution-based temporal attention mechanism that explicitly encodes temporal ordering and assigns higher weights to temporally closer items. To overcome GNN noise aggregation, we design a single-layer Spatio-Temporal Attention Graph Convolutional Layer (STA-GCL) that achieves higher-order neighbor aggregation while avoiding over-smoothing. For computational efficiency, we replace quadratic self-attention with a linear-complexity Mamba mechanism that selectively filters noise while efficiently capturing long-term dependencies. Additionally, we introduce virtual nodes and a global-level sparse attention encoder to further mitigate noise and capture inter-session dependencies. Extensive experiments on two public datasets demonstrate that MSTA-GNN significantly outperforms state-of-the-art methods, achieving improvements of 10.22%-14.89% in precision metrics and 1.1%-7.66% in MRR metrics.