Existing Transformer-based shadow removal methods are limited by fixed window sizes, making it difficult to effectively model global information. In addition, they do not fully utilize the distance prior in shadow images. This study argues that shadow removal should model brightness variations between regions from a global perspective. Non-shadow areas near the shadow boundaries are the most important for restoring brightness in shadow regions, and their importance gradually decreases as the distance increases. To achieve this, a regional decay attention mechanism is proposed, which introduces a positional decay bias into the self-attention computation to enable dynamic modeling of contributions from different spatial positions. A local perception module is introduced to improve the model’s ability to capture local details, and a shadow removal model named FW-Former is developed. This model achieves superior performance across multiple datasets, demonstrates stable generalization capability, and maintains a low parameter count.

Accurate segmentation of brain tumors from multimodal magnetic resonance imaging (MRI) is essential for diagnosis, treatment planning, radiotherapy targeting, and longitudinal assessment. Deep learning has advanced this task through convolutional neural networks, Transformers, state-space models, diffusion methods, and foundation models. However, strong benchmark performance does not guarantee clinical reliability because systems remain vulnerable to missing or degraded modalities, cross-center shift, and subregion-specific failures, particularly in enhancing tumor (ET) and tumor core (TC). This survey critically reviews deep learning for multimodal brain tumor segmentation from a deployment-oriented, failure-focused perspective. We introduce a unifying framework linking input reliability, fusion-architecture co-design, subregion-specific failure mechanisms, and uncertainty-aware clinical triage. Using this framework, we compare major architectural paradigms in contextual modeling, boundary preservation, computational feasibility, and robustness across ET, TC, and whole tumor (WT). We reinterpret multimodal fusion as a reliability-allocation problem and examine early, intermediate, token-level, sequence-aware, and adaptive strategies under incomplete or degraded inputs. We also synthesize robust learning approaches for sparse supervision, MRI quality degradation, cross-center variation, and test-time adaptation, and assess interpretability, uncertainty estimation, and human-in-the-loop review as mechanisms for clinical risk control. Finally, our evidence-oriented benchmarking analysis identifies WT performance saturation, persistent ET/TC instability, inconsistent boundary-metric reporting, and insufficient stress testing, center-stratified evaluation, calibration assessment, and computational transparency. We conclude that progress should be judged not only by benchmark accuracy but also by subregion-level reliability under realistic deployment conditions.

The proliferation of Internet of Things (IoT) devices makes them critical infrastructure, yet their scale and openness expose them to increasing cyberattacks. Existing dynamic IoT Intrusion Detection Systems (IDSs) can partially prevent cyberattacks. However, they typically demand extensive, high-quality labeled data, which makes updates costly and labor-intensive. Recently proposed Large-Language-Model (LLM)-based IoT IDSs address data labeling. However, they often over-rely on end-to-end LLM processing instead of hybrid LLM-Machine Learning (ML) approaches, which results in a significant detection delay. Meanwhile, the IDS’s data security is seldom considered in these works. This paper introduces the LLM and Post-Quantum-Secure Blockchain (QBc)-based adaptive Intrusion Detection System (LQB-IDS). LQB-IDS applies an adaptive detection method. It utilizes a lightweight ML model to detect known attacks that previously appeared in the training dataset. Meanwhile, all unknown attacks are handled by the LLM for dynamic analysis, using external knowledge secured by LLM managed QBc. Thus, it can continuously detect and adapt for unknown attacks without relying on pre-labeled data. Its core contributions include 1) an LLM-based unknown IoT traffic analyzer eliminating manual labeling; 2) a dual-switch learning approach combining LLM with lightweight ML models for low detection delay; and 3) a novel credit scoring algorithm with an LLM-Maintained Post-Quantum-Secure Blockchain-Based Database (LLM-QBc-DB) ensuring data security. Experimental results demonstrate that LQB-IDS achieves a 0.98 detection rate and a 0.97 average F1-score without labeled data, presenting an efficient and scalable solution for emerging IoT threats.

To address the increasing threats targeting Internet of Things (IoT) networks, the development of IoT Intrusion Detection Systems (IDS) has accelerated in recent years. However, existing IDSs often rely on manually labeled data and lack explainability, limiting their adaptability and practical deployment. This paper proposes an Explainable and Adaptive Internet of Things Intrusion Detection System (EADL-IDS) that combines a Denoising Convolutional Autoencoder (DCAE) for unsupervised binary anomaly detection with Large Language Models (LLMs) for multi-class attack interpretation via Retrieval-Augmented Generation (RAG). Compared to prior machine learning and deep learning based IDSs, EADL-IDS eliminates the need for labeled data while achieving higher detection accuracy. Unlike conventional LLM-based IDSs, it integrates a more effective Deep Neural Network (DNN) backbone and structured knowledge retrieval to improve explainability and reduce inference delay. Experiments on the Network Security Laboratory - Knowledge Discovery and Data Mining (NSL-KDD), Canadian Institute for Cybersecurity Internet of Things 2023 (CIC-IoT-2023), and Army Cyber Institute Internet of Things Network Traffic Dataset 2023 (ACI-IoT-2023) datasets show that EADL-IDS achieves over 95% binary accuracy and over 85% multi-class accuracy without supervision, offering a scalable and robust solution for IoT security.

Conventional multimodal perception systems for autonomous driving rely on data-driven learning and lack explicit mechanisms for reasoning about domain knowledge, such as sensor reliability and temporal risk, a gap that matters most where modality reliability varies, as in tunnel transitions and nighttime driving. We propose Knowledge-Aware Modality Enhancement (KAME), a framework that formalises sensor physics and driving-domain knowledge as executable rules and injects them into Multimodal Large Language Models (MLLM) as structured natural-language context. The RGB frame enters only as reliability metadata rather than as an encoder input, so the two modalities combine exclusively in the model's language-level reasoning. We call this knowledge-level frame-event fusion, as distinct from fusion at the data, feature, token, or decision level, and it leaves the model architecture unchanged. KAME comprises a Modality-Aware Knowledge Descriptor (MAKD) that encodes sensor reliability from physical measurements and a Temporal Dangerous Event Knowledge Extractor (TDKE) that models safety-critical temporal dynamics in event streams, both generated deterministically so that the reasoning remains interpretable and traceable. On 500 samples from the Driving Stereo Event Camera dataset, KAME raises the average Perceptual Accuracy of EventGPT-7B from 0.68 to 0.80 without retraining, while an adapted training-based feature-fusion baseline reaches 0.73, confirming that knowledge injection and feature alignment address different failure modes. The gains also transfer to a general-purpose MLLM and partially generalise to MVSEC, where 50-sample recalibration further improves performance. These results establish KAME as a portable, interpretable, and training-free approach to sensor-physics-aware MLLM perception under illumination degradation.

Diabetic retinopathy (DR) remains a major cause of visual impairment globally, and early, accurate diagnosis is critical for effective intervention. To address the challenges of limited labeled training data, difficulty identifying subtle and dispersed lesions, and redundant feature extraction in retinal images, this paper proposes a DR classification network with a multi-frequency contextual attention module (MFCA-DRNet). First, this paper introduces a self-supervised contrastive learning strategy pre-trained on the EyePACS dataset, eliminating reliance on extensive labeled data and enabling effective feature learning. Next, an adaptive preprocessing method, integrating histogram equalization with non-local means denoising, is designed to enhance image quality by reducing noise and improving lesion visibility. The proposed MFCA module effectively captures long-range contextual relationships and associates dispersed lesion characteristics across retinal images, significantly enhancing lesion recognition. Additionally, the backbone network incorporates an attention mechanism guided by an energy function, emphasizing lesion-specific features while suppressing irrelevant information. Evaluated through downstream classification tasks on DDR, APTOS 2019, and Messidor-2 datasets, MFCA-DRNet achieved strong performance, particularly on the APTOS 2019 dataset, with accuracy, precision, recall, andF1 scores of 87.12%, 81.2%, 85.3%, and 83.16%, respectively. These results highlight MFCA-DRNet's potential to improve DR diagnosis and clinical applicability in diverse imaging conditions.

Image shadow removal is a typical low-level vision task, as shadows introduce abrupt local brightness variations that degrade the performance of downstream tasks. Due to the quadratic complexity of Transformers, many existing methods adopt local attention to balance accuracy and efficiency. However, restricting attention to local windows prevents true long-range dependency modeling and limits shadow removal performance. Recently, Mamba has shown strong ability in vision tasks by achieving global modeling with linear complexity. Despite this advantage, existing scanning mechanisms in the Mamba architecture are not suitable for shadow removal because they ignore the semantic continuity within the same region. To address this, a boundaryregion selective scanning mechanism is proposed that captures local details while enhancing continuity among semantically related pixels, effectively improving shadow removal performance. In addition, a shadow mask denoising preprocessing method is introduced to improve the accuracy of the scanning mechanism and further enhance the data quality. Based on this, this paper presents ShadowMamba, the first Mambabased model for shadow removal. Experimental results show that the proposed method outperforms existing mainstream approaches on the AISTD, ISTD, SRD, and WSRD+ datasets, and demonstrates good generalization ability in cross-dataset testing on USR and SBU. Meanwhile, the model also has significant advantages in parameter efficiency and computational complexity. Code is available at: https://github.com/ZHUXIUJINChris/ ShadowMamba.

Vehicle Make and Model Recognition (VMMR) is a fine-grained classification task in Intelligent Transportation System (ITS). Recent works address VMMR through feature encoding schemes, part-based methods or attention modules. Despite their astounding results, these techniques concentrate on the high-level semantic features. This practice cripples the feature expressive ability of the networks as the granular traits of the vehicle distilled from the early convolution layers are not embedded into the final feature representations. In this work, by contrast, a Cross-Granularity (CG) module which is responsible for the integration of macroscopic and microscopic components is proposed. By incorporating the CG module into a Convolutional Neural Network (CNN), the resultant network i.e. CGNet reinforces the feature extraction ability by amalgamating the feature maps from different scales to render a balanced mix between local contextual information and global semantic details. To validate the proposed framework, experiments are conducted on four publicly available datasets. We report competitive performance on web-nature Comprehensive Cars, Stanford Cars, Car-FG3K and surveillance-nature Comprehensive Cars datasets with 98.3%, 95.4% 86.4% and 99.1% accuracies. Furthermore, we demonstrate the ability of the CGNet to pinpoint distinctive fine-grained details via the Gradient-Weighted Class Activation Mapping (Grad-CAM) technique and compare it against the baseline which learns on deep features alone. The generalization ability of the CG module on other CNNs is also examined and the results suggest a high compatibility between the two.

Accurate State of Health (SOH) prediction is vital for ensuring reliable energy supply and timely battery replacement in battery management systems (BMS). While most studies focus on SOH estimation under constant discharge conditions, the impact of varying currents remains underexplored. This work introduces a novel health indicator (HI), Half Load Discharging Time (HLDT), to model battery aging by measuring the time required to reach half capacity under fluctuating currents. Additionally, a Quantum-Enhanced Ensemble Model (QEM) is proposed, integrating quantum computing principles—superposition and entanglement—with classical models to enhance prediction accuracy. The model combines Quantum Support Vector Regression (QSVR) and Gated Recurrent Units (GRU) in parallel, where QSVR leverages quantum kernels for feature mapping, and GRU captures temporal dependencies. The outputs from both QSVR and GRU are integrated into Gaussian Process Regression (GPR) to enhance prediction refinement. The QEM achieves an average prediction error of 1.11% and a correlation coefficient of 0.998 across four NASA datasets, outperforming conventional models. Through SHapley Additive exPlanations (SHAP) analysis, voltage kurtosis has been identified as a critical battery HI with a significant impact on SOH prediction.

Shadow removal is essential for improving image visual quality, as shadows often block important details and cause color distortion. At present, most shadow removal models still have clear limitations in handling this task. On the one hand, many methods rely solely on a single attention mechanism, which makes it difficult to simultaneously achieve illumination transfer, local detail restoration, and effective selection of key channels. On the other hand, even with a multistage design, many models lack real-time correction between stages, making it hard to fix errors in time and affecting the overall stability and accuracy of the results. In addition, existing algorithms often ignore the background prior information available in shadow images. Invariant images usually contain stable illumination-invariant features, which help suppress gradient changes around shadow boundaries and enhance background representation. To address these issues, this study proposes a Multistage Progressive Invariant Image-guided Shadow Removal Network (MPISR-Net). The network divides the complex shadow removal task into three stages: initial feature extraction, illumination transfer, and channel refinement. Real-time supervision is introduced at each stage to progressively remove shadows. Moreover, invariant image features are integrated between stages and used to guide the network’s focus on background information, effectively reducing shadow interference. Extensive experimental results show that the proposed method achieves excellent performance on several commonly used shadow removal datasets, and effectively handles complex shadows through multistage processing.

Introduction: Accurate multi-pollutant forecasting is vital for urban governance and public health. Existing deep models struggle to capture multi-scale temporal dynamics and synergistic cross-pollutant relations. Methods: We propose an Enhanced Bidirectional Attention Multi-scale Temporal Network (EBAMTN) that combines a multi-scale TCN with linear attention, a two-layer BiLSTM augmented by multi-head self-attention, and a gated fusion layer. Under a multi-task paradigm, the backbone jointly learns shared temporal representations and outputs PM2.5 and PM10 via task-specific heads. Results: Using hourly data from Guangzhou, Beijing, and Chengdu, EBAMTN achieved R2 > 0.94 for both pollutants while maintaining low errors (e.g., PM2.5 MAE approximate to 2.03, RMSE approximate to 2.94; PM10 MAE approximate to 3.44, RMSE approximate to 4.99). Confidence-interval analyses and scatter plots indicate strong trend tracking and robustness, with remaining challenges mainly at sharp peaks. Discussion: The integration of multi-scale convolutions, bidirectional memory, attention, and gated fusion improves accuracy, interpretability, and generalization. The lightweight design (approximate to 2.1M parameters; similar to 13.2 ms/sample) supports real-time and edge deployment. Overall, EBAMTN offers a scalable, interpretable solution for multi-pollutant forecasting in complex urban settings.

Accurate air quality forecasting is crucial for public health and environmental policy, yet existing deep learning models often suffer from high computational costs and inadequate modeling of temporal dependencies and multi-pollutant interactions. This paper proposes a novel deep learning framework that integrates an Enhanced Transformer with Bidirectional Mamba2, optimized through multi-task learning and knowledge distillation. Using a teacher–student paradigm, the teacher model captures rich temporal semantics, while the lightweight student model retains predictive accuracy with reduced inference costs. Key innovations include a hybrid architecture combining multi-scale global-local attention and long-range dependency modeling, a regression-specific knowledge distillation approach with soft target smoothing and intermediate representation transfer, and an end-to-end multi-task design for joint forecasting of multiple pollutants. Extensive experiments on real-world datasets from Guangzhou, Chengdu, and Beijing (2018–2022) across four key pollutants (PM2.5, PM10, NO2, SO2) demonstrate that our model significantly outperforms baseline methods, with the student model maintaining accuracy within 5% of the teacher while requiring fewer parameters. These results highlight the framework’s potential for accurate, scalable, and efficient air quality forecasting in real-time and resource-constrained environments.

Image-text matching is a vital task in multi-modal intelligence. Recently, researchers have moved beyond simply aligning fragments between image regions and text words at a low level. They are now focusing on high-level alignment with semantic relations to enhance performance, and significant progress has been made. However, almost all existing research concentrates on extracting semantic relations between fragments within individual modalities, overlooking the interaction between modalities. This oversight results in a deficiency of high-level semantic and consistency information, ultimately diminishing performance. To address this, we argue that integrating intra-modal and inter-modal semantic relations between fragments could help improve feature quality and promote consistency in semantic space by enabling more comprehensive interactions between modalities. Therefore, we propose a novel Cross-modal Semantic Relations Enhancement Network (CSREN), which designs an implicit semantic relations graph and employs a graph attention network to adaptively mine semantic relations from both modalities, separately and collectively. Quantitative experiments on Flickr30K and MS-COCO demonstrate that our method significantly outperforms existing approaches, achieving improvements of 5.4% to 24.5% for text retrieval and 9.7% to 26.9% for image retrieval.

The removal of shadows from images is a classic problem in computer vision, aiming to restore the lighting in shadowed areas, thereby reducing the information interference and loss caused by the presence of shadows. In recent years, numerous excellent shadow removal algorithms have emerged, particularly with the rapid development of deep learning technology, which has disrupted traditional physics-based approaches and significantly improved the effectiveness of shadow removal. In this paper, we conduct a comprehensive survey of shadow removal methods published from 2017 to the present. We first introduce background knowledge about image shadow removal, providing detailed explanations of both physics-based and learning-based shadow removal methods. We analyze and compare these algorithms from both quantitative and qualitative perspectives, reassessing all models that provided open-source result sets according to uniform criteria. Additionally, we introduce commonly used datasets and evaluation metrics in the field. Finally, we discuss applications of shadow removal in specific scenarios, along with research challenges and opportunities in this domain.

Alzheimer's disease (AD) is a neurodegenerative ailment that is becoming increasingly common, making it a major worldwide health concern. Effective care depends on an early and correct diagnosis, but traditional diagnostic techniques are frequently constrained by subjectivity and expensive costs. This study proposes a novel Vision Transformer-equipped Convolutional Neural Networks (VECNN) that uses three-dimensional magnetic resonance imaging to improve diagnosis accuracy. Utilizing the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset, which comprised 2,248 3D MRI images and diverse patient demographics, the proposed model achieved an accuracy of 92.14%, a precision of 86.84%, a sensitivity of 93.27%, and a specificity of 89.95% in distinguishing between AD, healthy controls (HC), and moderate cognitive impairment (MCI). The findings suggest that VECNN can be a valuable tool in clinical settings, providing a non-invasive, cost-effective, and objective diagnostic technique. This research opens the door for future advancements in early diagnosis and personalized therapy for Alzheimer's Disease.

Worldwide, cardiovascular disease is the leading cause of death. Based on clinical data, a Machine Learning (ML) system can detect cardiac disease in its early stages, which enables a reduction in mortality rates. However, imbalanced and high dimensionality data have been a persistent challenge in ML, impeding accurate predictive data analysis in many real-world applications, such as the detection of cardiovascular disease. To address this, computational methods targeting heart disease detection have been developed. However, their performance is still inadequate. Hence, this study presents a new stack predictor for the heart disease model (termed SPFHD). SPFHD employs five common tree-based ensemble learning algorithms as base models for heart disease detection. In addition, the predictions from the base models are integrated using a support vector machine algorithm to enhance the accuracy of heart disease detection. A new conditional variational autoencoder (CVAE) based method is developed to overcome the imbalance issue, which performs better than the conventional balancing methods. Finally, the SPFHD model is tuned by Bayesian optimization. The results show that the proposed SPFHD model outperforms the state-of-art methods over four datasets achieving higher f1-score of 4.68 %, 4.55 %, 2 %, and 1 % for HD clinical, Z-Alizadeh Sani, Statlog, and Cleveland, respectively. Moreover, this new framework offers vital interpretations which assist in understanding model success by leveraging the powerful SHapley Additive explanation (SHAP) algorithm. This highlights the most significant attributes for detecting heart disease and overcoming the limitations of current 'Black-box' methods that cannot reveal causal relationships between features.

Mobile Edge Computing (MEC) has paved the way for new Cellular Internet of Things (CIoT) paradigm, where resource constrained CIoT Devices (CDs) can offload tasks to a computing server located at either a Base Station (BS) or an edge node. For CDs moving in high speed, seamless mobility is crucial during the MEC service migration from one base station (BS) to another. In this paper, we investigate the problem of joint power allocation and Handover (HO) management in a MEC network with a Deep Reinforcement Learning (DRL) approach. To handle the hybrid action space (continuous: power allocation and discrete: HO decision), we leverage Parameterized Deep Q-Network (P-DQN) to learn the near-optimal solution. Simulation results illustrate that the proposed algorithm (P-DQN) outperforms the conventional approaches, such as the nearest BS +random power and random BS +random power, in terms of reward, HO cost, and total power consumption. According to simulation results, HO occurs almost in the edge point of two BS, which means the HO is almost perfectly managed. In addition, the total power consumption is around 0.151 watts in P-DQN while it is about 0.75 watts in nearest BS +random power and random BS +random power.

Alzheimer's disease (AD) is an illness that involves a gradual and irreversible degeneration of the brain. It is crucial to establish a precise diagnosis of AD early on in order to enable prompt therapies and prevent further deterioration. Researchers are currently focusing increasing attention on investigating the potential of machine learning techniques to simplify the automated diagnosis of AD using neuroimaging. The present study involved a comparison of models for the detection of AD through the utilization of 2D image slices obtained from magnetic resonance imaging brain scans. Five models, namely ResNet, ConvNeXt, CaiT, Swin Transformer, and CVT, were implemented to learn features and classify AD based on various perspectives of 2D image slices. A series of experiments were conducted using the dataset from the Alzheimer's Disease Neuroimaging Initiative. The results showed that ConvNeXt outperformed ResNet, CaiT, Swin Transformer, and CVT. ConvNeXt exhibited an average accuracy, precision, recall, and F1 score of 95.74%, 96.71%, 95.74%, and 96.14%, respectively, when applied to a 3-way classification task involving AD, mild cognitive impairment, and normal control subjects. The results suggest that the utilization of ConvNeXt may have potential in the identification of AD using 2D slice images.

Digitalization is revolutionizing our way of life and catalyzing the transformation into smart city. Intelligent Transportation System (ITS) being an indispensable component of smart city leverages massive amount of collected information to improve traffic efficiency, thereby creating a safer and comfortable commuting environment for the users. One of the most important tasks in ITS is vehicle classification which aims to find out the vehicle identity, including vehicle segment, automobile maker, model, etc. In this article, we first present the vehicle classification taxonomy branched based on the nature of input data. We subsequently investigate diverse area of sensor-based vehicle classification followed by image-based vehicle classification which cover both the conventional and emerging techniques in a comprehensive manner. The methodologies together with the corresponding strengths and potential weaknesses are elucidated so that it serves as an invaluable reference for vehicle classification related applications in the future. More importantly, we express our views on future research direction with the intention to accelerate the development of vehicle classification field. In contrast to previous works, we aim to cover wide spectrum of vehicle classification methodologies in this review to provide more clarity when it comes to selecting a solution that suits individual need. They include both sensor-based and image-based vehicle classification for VTR, VLR and VMMR. We employ exhaustive coverage approach for the former to identify the extant solutions that are built upon various kinds of sensing technologies and they are further collated according to the installation methods. For image-based methods, we screen for the works that have been central to both pre- and post-deep learning era. The featured works either address the shortcoming of previous works in image domain or present novel concept to advance the classification performance.

Distracted drivers are more likely to get involved in a fatal accident. Thus, detecting actions that may led to distraction should be prioritized to reduce road accidents. However, there are many actions that cause a driver to pivot his attention away from the road. Previous works on detecting distracted drivers are done through a defined set of actions that are considered as distraction. This type of dataset is known as ‘closed set’ since there are still many distraction actions that were not considered by the model. Being different from previous datasets and approaches, in this work, we utilize constructive learning to detect distractions through multiview and multimodal video. The dataset used is the Driver Anomaly Detection dataset. The model is tasked to identify normal and anomalous driving condition in an ‘open set’ manner, where there are unseen anomalous driving condition in the test set. We use Video Transformer as the backbone of the model and validate that the performance is better than convolutional-based backbone. Two views (front and top) of driving clips on two modalities (IR and depth) are used to train individual model. The results of different views and modalities are subsequently fused together. Our method achieves 0.9892 AUC and 97.02% accuracy with Swin-Tiny when considering both views and modalities.