Applied Intelligence最新文献

Object detection, a fundamental task in computer vision, has been extensively employed in numerous machine learning contexts. Nevertheless, object detectors are susceptible to various attacks and present significant security concerns in practical applications. As a particularly insidious attack, the backdoor attack involves embedding a hidden backdoor into the object detector, which can lead to misleading results. However, the majority of existing research on backdoor attacks employs a single pattern in the spatial domain of image as a trigger, which inevitably destroys the pixel-level semantics of benign image. To address this, we propose a novel Backdoor Attack against Object Detection via Frequency Noise Injection, i.e., FBAO. We employ the Gaussian random noise function to generate a noise image, which is then injected into the benign image by linearly combining the amplitude spectra of the perturbation and the benign image. By preserving the pixel-level semantics of benign images when injecting triggers, FBAO ensures the invisibility of generated triggers. Furthermore, we design the object-based evaluation of the Object-based Attack Success Rate (OASR) and the Object-based Miss-triggering Rate (OMR), which introduce the prediction of bounding box to comprehensively assess the effectiveness of backdoor attack against object detection. Experimental results show consistent out-performance of our method over other baselines across different object detection models and datasets.

Cold start is a challenge in the field of recommender systems. Cross-domain Recommendation (CDR), as a promising approach to solve this problem, focuses on how to effectively transfer user preferences from one domain to another. Current approaches have the following issues: user’s interests extraction depending on priori knowledge; noise being easily introduced and negative transfer arising when user’s features are transferred; users’ overall features and diverse interests being less considered together. This paper proposes a novel cross-domain recommendation method named Cross-domain Recommendation based on Three-Phase Cross-Attention and Multi-Granularity Transfer Meta-Network (CAMGCDR). The self-attention mechanism is adopted to extract the interest features from the user’s behavioral sequences, so that there is no more relying on priori knowledge. The item-level cross-attention is used to fine-tune the embedding vectors of the user’s historical behavior in the source domain. The part of the user’s historical behaviors in the source domain that are most relevant to the item in target domain are paid attention, resulting in less noise and negative transfer. By constructing mapping bridges with two granularities, the user embedding vectors and multi-interest vectors are mapped to target domain respectively. User embedding vector is concatenated with multi-interest vectors so that both overall features of the user and the diverse interests are taken into account. Thus, the accuracy and personalization recommendation of CAMGCDR are improved. Experimental results on large-scale real-world datasets show that the CAMGCDR outperforms all the baseline methods, proving its effectiveness and practicality.

Traditional Chinese medicine (TCM) electronic medical record (EMR) is the core carrier to record the patient’s diagnosis and treatment process and TCM discursive thinking, and its Nested Named Entity Recognition (Nested NER) is of great significance for constructing TCM knowledge graph and intelligent application. However, existing methods face three major challenges: insufficient model generalization ability due to the scarcity of TCM text annotation data, catastrophic forgetting problem in domain adaptation of large language models, and semantic illusion phenomenon of generative models in nested entity recognition. To address the above problems, this study proposes a three-stage optimization framework approach, KDC-NER: first, multi-dimensional data augmentation of self-constructed TCM electronic medical record dataset based on an improved EDA method, and expanding the domain corpus through strategies such as synonym substitution, entity replacement, and so on; second, designing a dynamic data filtering mechanism for knowledge augmentation in TCM, combining the entity distribution a priori with semantic similarity computation, to alleviate the big model knowledge forgetting problem in fine-tuning; finally, a constraint generation method based on cue engineering is proposed to suppress non-relevant entity illusions in the generation process through entity boundary-aware templates and knowledge verification modules. The experimental results show that the proposed method achieves 80.57% F1 value in the self-constructed TCM electronic medical record dataset, which is improved compared with the traditional BERT-BiLSTM-CRF, GP, and GPT-3.5-turbo baseline models, and verifies the effectiveness of the data augmentation, dynamic filtering and constraint generation strategies. This study provides a scalable solution for nested entity recognition in the TCM domain and offers new ideas for large model domain adaptation research in low-resource scenarios.

The widespread use of digital images for sharing and communication has led to increased focus on the protection of digital rights, with watermarking serving as a critical tool for safeguarding these assets. While traditional watermarking algorithms are effective, they often face challenges with robustness under various distortions. The integration of deep learning techniques, such as Hopfield Neural Networks (HNN), presents a promising approach to enhancing both the robustness and security of watermarking systems. However, achieving a balance between robustness and security in adversarial conditions remains a challenge. To address these issues, this study proposes a novel digital image watermarking algorithm that combines a two-dimensional discrete memristor chaotic map (QDM-CM), which we developed for watermark encryption, with a Hopfield Neural Network for post-processing error correction. Specifically, the proposed algorithm employs chaotic sequences for watermark encryption and utilizes DCT to embed the encrypted watermark image into the host image. Additionally, the innovative application of the HNN ensures effective recovery of the watermark under a range of attacks. Through extensive simulations, we demonstrate that the proposed method provides high robustness against common attacks, such as rotation, low-pass filtering, shear, noise interference, and JPEG compression, outperforming conventional watermarking schemes. These findings underscore the potential of combining chaotic encryption and associative memory to strengthen digital watermarking in real-world applications.

With the rapid development of the Internet, multimodal information has become increasingly abundant and structurally complex, making multimodal recommendation systems critical in practical applications. In recent years, substantial efforts have been devoted to alleviating data sparsity and cold start issues in recommendation systems to improve their adaptability in complex multimodal recommendation scenarios. However, the existing methods often ignore the uniqueness of noise between different modes when dealing with modal noise, and fail to achieve targeted denoising. During the modality fusion process, the behavioral differences across different modalities are not fully considered, which limits the recommendation performance of the model. Therefore, we propose a novel Spectrum-guided Denoising and Two-Stage Fusion for Multimodal Recommendation framework (SDFMRec). Specifically, we leverage spectral transformation to perform independent denoising for each modality, effectively mitigating intermodal noise interference. Furthermore, we introduce a dual-stage fusion strategy that jointly captures global semantic correlations and local behavioral differences between modalities, improving both the precision and robustness of recommendations. Extensive experiments on four public multimodal recommendation datasets demonstrate that our method significantly outperforms state-of-the-art baselines, validating its effectiveness and superiority.

Tools for automatic wood species identification are needed worldwide in order to support sustainable timber trade. This work explores the application of computer vision techniques to classify high-resolution macroscopic images of timber. The main challenge of this problem is that fine-grained patterns in timber are crucial in order to accurately identify wood species, and these patterns are not learned by convolutional neural networks (CNNs) trained on low resolution images. This work introduces the Timber Deep Learning Identification with Patch-based Inference Voting methodology, abbreviated TDLI-PIV methodology. This methodology exploits the concept of patching and the availability of high-resolution macroscopic images of timber in order to overcome the inherent challenges that CNNs face in timber identification. The TDLI-PIV methodology is able to capture fine-grained patterns in timber and, moreover, boosts robustness and prediction accuracy via a collaborative voting inference process. In this work we also introduce a new data set of marcroscopic images of timber, called GOIMAI-Phase-I, which has been obtained using optical magnification, and is openly published online in zenodo. Our experiments have assessed the performance of the TDLI-PIV methodology, including a comparison with other methodologies available in the literature, an exploration of data augmentation methods and the effect that the dataset size has on the accuracy of TDLI-PIV.

Crowd evacuation simulations provide guidance for emergency response to personnel emergencies in public places. Existing path planning methods fail to account for the impact of crowd panic, thus reducing the crowd evacuation in reality. To address this issue, we propose a panic emotion aware path planning method for crowd evacuation. First, the ResNet-RP (Residual Neural Network - Recognition Panic) model with spatiotemporal similarity constraints accurately identifies individual panic emotions, and calculates the panic level of the crowd based on the degree of individual aggregation. Second, the proportion of panicked people is predicted via the mean-field method on the basis of the panic degree of the crowd. Finally, the predicted proportion of panicked people is introduced into the reward function of the multi-agent deep deterministic policy gradient algorithm (P-MAD) to realize emotion-aware evacuation path planning. The experimental results show that our proposed method can effectively achieve panic emotion awareness and panic avoidance path planning for crowd evacuation.

In the realm of online education, it is a fundamental task to assess the mastery of knowledge points by students and to customize personalized exercises for them. However, recommending suitable exercises for students poses a challenge due to the varying levels of student knowledge and the extensive exercise question bank. Current methodologies utilize collaborative filtering to recommend exercises, which lack consideration of changes in student knowledge, making it difficult to capture student behavior and utilize the relationship information between students and exercises. Therefore, this paper proposes an end-to-end knowledge graph (KG) enhanced exercise recommendation algorithm for multi-task learning (KERM). By incorporating a feature sharing unit, the recommendation task and knowledge graph embedding (KGE) task automatically share additional feature information with each other, thereby enhancing the accuracy of exercise recommendation. Firstly, a self-adaptive adjustment factor has been designed that continuously monitors the changes in knowledge state of the students in the recommendation task. Subsequently, a cleverly designed single multi-level feature interaction is implemented in the KGE task where single-level feature interactions extract detailed entity information and capture rich interaction details within entity neighborhoods, multi-level feature interactions aggregate multiple single-level interactions to expand receptive field coverage and improve quality of feature interaction. Finally, a feature sharing unit is designed to model high-order interactions between student and exercise features by automatically sharing additional information between both tasks to help prevent overfitting while improving robustness. To verify the effectiveness of KERM, extensive experiments are conducted on four real datasets yielding average prediction accuracy by 4.2%, 5.6%, 4.1%, and 4.9% respectively.

Locusts are a notorious pest that severely impacts agricultural production in China. Efficient segmentation and detection of camouflaged locusts are essential. However, previous segmentation methods often exhibit limitations in handling complex backgrounds or struggles with fine detail processing, particularly with small features like locust antennae. To address this problem, we propose an EW-CASCADE model and introduce a new dataset, Camouflaged Locust. We design an Enhanced Efficient Multi-scale Attention (EEMA) module that leverages parallel sub-networks and spatial learning to avoid dimensionality reduction, thereby preserving additional feature information while enhancing the model’s ability to capture long-range dependencies. Meanwhile, we introduce Wavelet Transform Convolution (WTConv) and incorporate it to decompose information effectively across multiple scales, thus improving the decoder’s global perception as well as the edge and texture details. Experimental results demonstrate that the proposed EW-CASCADE model achieves superior segmentation accuracy on the Camouflaged Locust dataset compared to the previous state-of-the-art approaches, particularly in capturing fine details and low-contrast regions. Our model improves by 0.057 in mDic, 0.079 in mIoU, 0.071 in wFm, and 0.040 in Sm, while the MAE decreases from 0.022 to 0.013.

Real-time, high-quality computer-generated holography (CGH) is essential for next-generation virtual reality or augmented reality displays. Existing deep learning methods often struggle to balance computational efficiency with reconstruction fidelity. Complex-valued convolutional neural networks (CCNNs) are well-suited for phase-only hologram generation. However, achieving real-time, high-fidelity holograms remains challenging due to three key issues: insufficient modeling of complex-field feature interactions, phase distortion during up-sampling, and a lack of integration with physical optics principles. This study proposes an end-to-end physics-aware complex-valued attention network (PACAN-CGH) to address these challenges. Its core innovations include a complex-valued hybrid attention module for adaptive feature selection, a phase-continuous up-sampling layer based on complex-valued sub-pixel convolution, and a physics-driven loss function incorporating band-limited diffraction constraints. This co-design ensures high-quality hologram generation that is both computational efficiency and physical consistency. Experiments validate the superiority of PACAN-CGH, achieving an average Peak Signal-to-Noise Ratio of 33.31 dB at 1920(times)1072 resolution with fast inference time of 0.36 seconds per frame. Ablation studies confirm the contribution of each component, and cross-dataset tests demonstrate superior generalization capability. This work bridges optical physics with neural network design, establishing a new paradigm for efficient and physically interpretable CGH, and advancing complex-valued neural network design.