Prioritized Experience Replay (PER) is a technical means of off-policy deep reinforcement learning by selecting important experience samples more frequently to improve the training efficiency. However, the non-uniform sampling applied in PER inevitably shifts the state-action distribution and brings the estimation errors of Q-value function. In this paper, an efficient off-policy reinforcement learning training framework called Attention Loss Adjusted Prioritized (ALAP) Experience Replay, is proposed. ALAP exploits the similarity of the transitions in buffer to quantify the training progress, and accurately corrects the bias based on the positive correlation between the error compensation strength and the training progress. In order to verify the effectiveness of the algorithm, the ALAP is tested on 15 different games on Atari 2600 benchmark. Additionally, we developed a multi-USV competition scenario using Unreal Engine to further illustrate the superiority as well as the practical value of ALAP.
The session-based recommendation provides accurate recommendations for anonymous users by analyzing their preferences through their behavior data in recent times. Existing session-based recommendation systems methods do not adequately utilize the click orders and click counts of session items. They do not fully capture the dependency between each item in a session. We propose a new model for session-based recommendation, RESR-GNN, to resolve these issues. The main contributions of the paper are as follows: (1) To differentiate the importance of different items, designing the corresponding relationship-enhanced session graph with the session, including click information; (2) Soft attention and multi-layer self-attention mechanisms are used to comprehensively analyse preferences between each item in an entire session. We conducted multiple comparative experiments using two publicly available datasets, and our findings show that RESR-GNN performs better than other baseline models.
With the development of intelligent manufacturing, higher requirements for real-time performance and accuracy have been placed on the inspection of surface quality in industrial products. As a core basic material in manufacturing, steel has various complex surface defects, such as cracks, scratches, and scale, which are characterised by small size, diverse shapes, and strong background interference, posing significant challenges for automated inspection. To address the issues of insufficient detection accuracy, limited feature fusion capabilities, and coupling of classification and localisation tasks in existing YOLO models when processing fine-grained defects on steel surfaces, this paper proposes a high-performance object detection algorithm based on an improved YOLOv8m: ACT-YOLO (Adaptive Content-guided Task-aligned YOLO). This algorithm integrates three key modules: the AFMA module to enhance multi-scale perception capabilities for small objects; the CGAF module to achieve content-guided multi-attention feature fusion; and the TADD module to optimise the dynamic alignment between classification and regression tasks in the detection head. Evaluations on the NEU-DET steel surface defect benchmark dataset demonstrate that ACT-YOLO achieves an mAP@0.5 of 86.4% and a detection speed of 115 FPS. Compared to non-YOLO methods such as SSD (mAP@0.5: 61.0%, FPS: 41), RetinaNet (mAP@0.5: 69.5%, FPS: 15), and RT-DETR-r101 (mAP@0.5: 78.8%, FPS: 108), as well as other YOLO series models, ACT-YOLO exhibits significant advantages in both detection accuracy and real-time performance. Generalisation experiments on the GC10-DET dataset also validate its cross-scenario adaptability. ACT-YOLO balances detection accuracy, speed, and model lightweighting, making it suitable for the demand for efficient, real-time defect detection systems in actual industrial environments, with broad engineering application prospects and research value.
The evaluation of regional pulmonary ventilation function is of significant clinical value, particularly in the initial diagnosis of pulmonary disorders, staging assessment, and personalized treatment planning. This study proposes a multi-scale VMamba attention registration network (MS-VMANet) to predict 4DCT pulmonary ventilation changes using unsupervised learning registration. MS-VMANet primarily integrates the efficient visual mamba attention, which captures long-range feature information globally, and the multi-head dilated regional attention improves deformation field prediction via aggregating multi-scale contextual features through dilated convolutions and attention mechanisms. Then, the deformation fields were calculated using the Jacobian determinant to generate images that reflect lung ventilation distribution to assess regional lung ventilation function. According to the experimental findings, the MS-VMANet performs better in terms of registration accuracy and performance, providing a reliable technical means for assessing regional pulmonary ventilation function.
The advantages of non-contact and embedded palm vein recognition technology in terms of security, privacy, and convenience are highlighted. Performance factors include error rate, rejection rate, and real-time processing, while engineering factors include small size, scalability, and cost-effectiveness. However, the non-contact sampling mode is susceptible to uncontrollable factors such as palm placement and posture, as well as external environmental influences like complex backgrounds and strong lighting. This paper proposes an omni-direction palm vein recognition (oPVR) method that integrates geometric and brightness multi-dimensional image quality feedback (MIQF), including integrity judgment, scale assessment, rotation angle detection. The method involves hand detection network, palm single-shot multibox detection (P-SSD) network for region of interest(ROI), and vein feature extraction network. These enhancements effectively improve recognition performance under environmental disturbances and pose variations. Based on a real-world database and a real-time demonstration system, a series of comparative tests and ablation experiments were conducted to validate the algorithm's accuracy, speed and usability for large-scale applications.
The paper deals with modification of super-twisting algorithm in sliding mode speed controller for predictive control of induction motor. At very-low-speed and/or high-load torque regions, the conventional proportional-integral controller gives poor speed response performance. In order to enhance the performance, sliding mode control (SMC) is deployed. However, the conventional SMC often owns high-chattering control signal. Super-twisting algorithm is utilized to reduce the chattering. In order to obtain high accuracy of the speed response, boundaries of the disturbance term and its time derivative are defined using dynamic equation of the motor. Simulations with the predict control of three controllers confirm the superior performance of the proposed super-twisting algorithm in terms of normalized integral time absolute error and relative final error of the motor speed.
Accurate prediction of drug-target binding affinity (DTA) is a core challenge in computer-aided drug design. In this study, we propose the TransGNN-DTA framework to construct a multi-scale feature representation and dynamic fusion mechanism by hierarchically integrating Transformer and graph neural network (GNN): 1) byte-pair encoding (BPE) based unified semantic characterization of drug SMILES and protein sequences, which preserves the atomic-level chemical structure and residue-level functional motifs; 2) hierarchical Transformer-GNN encoding architecture to capture sequence global dependencies (e.g., protein functional domain interactions) and molecular local structures (e.g., drug-functional group interactions); 3) chunked adaptive training strategy, hybrid accuracy, and the combination of an optimizer and scheduler to effectively reduce the hardware resource requirements for training. On the DAVIS and KIBA datasets, the model improves the CI metrics by an average of 3.23% over the existing optimal methods (DGraphDTA, GraphDTA). The computational efficiency is significantly optimized by systematic chunking, providing a scalable end-to-end solution for large-scale DTA prediction. The source code and datasets are publicly accessible at https://github.com/Quietpeng/TransGNN_DTA.
In naturalistic settings, object detection methodologies utilizing horizontal anchors exhibit commendable performance. However, when applied to remote sensing contexts, these methods may engender complications such as anchor misalignment and target-anchor overlap. This study addresses these challenges by introducing the S2A-AGC-Net, a rotating object enhancement network designed specifically for remote sensing applications. The S2A-AGC-Net integrates online data augmentation, group convolution, and lightweight CARAFE operators. For backbone feature extraction, an advanced strategy combining group convolution with ResNet101 enhances feature extraction capabilities. In the neck segment, the original FPN component is refined with the lightweight CARAFE upsampling operator, improving feature fusion and model speed. Additionally, a novel online data augmentation technique that combines Mosaic, Mixup, and HSV color transformations is introduced at the input to enhance the model’s generalization capacity. Ablation studies conducted on the HRSC2016 datasets reveal that the S2A-AGC-Net attains a mean Average Precision (mAP) value of 0.6048 while achieving 18.5 FPS, surpassing the benchmark S2A-Net by 3.38%. Performance evaluations on the more intricate DOTA datasets show a 1.39% improvement over the S2A-Net benchmark. Comparative analyses with existing state-of-the-art algorithms further corroborate the superior accuracy and efficiency of our proposed method. The findings underscore the effectiveness of integrating advanced augmentation techniques and efficient network architecture in improving detection outcomes. The progressive nature of the S2A-AGC-Net positions it as a promising solution for addressing the challenges of object detection in increasingly complex environments, paving the way for future research and development in remote sensing applications.
In recent years, a large amount of debris has appeared on power transmission lines, affecting circuit power supply and endangering human safety. At present, deep learning based object detection algorithms have made continuous progress, but they still cannot meet the practical application requirements of real-time performance. It is necessary to further reduce model complexity and improve detection speed. Therefore, a transmission line debris recognition based on fusion point attention improved YOLOV8 is proposed. The model uses the original YOLOV8 model as the base model and adds PixelAttention in detection to enhance the model's receptive field and multi-scale object perception, thereby improving the model detection accuracy. The experimental results show that the average accuracy of the original model is only 92.3%, while the average accuracy of the improved model has increased to 94%.
With the explosion of micro-videos, it is essential to develop video summariza-tion algorithms that simultaneously capture the diverse video content and represent theoriginal video. Many methods employ deep neural networks (DNNs). However, DNNsummarization models do not directly consider summary diversity. On the other hand,submodular functions can be seen as a form of diversity. However, the shallow structureprevents the data representation at a more abstract level. This paper proposes a novelsubmodular pseudo-3D network (SP3D), which equips submodular functions with a multi-layered network for micro-video summarization. Unlike standard DNNs, the proposedSP3D network and the correspondingoptimization method consider the interlock depen-dency among the selected frames, thus improving the diversity. The experimental resultsindicate that the proposed model and optimization method are effective in micro-videosummarization.
Triple negative breast cancer develops rapidly with high mortality and limited treatment. Therefore, it is important to detect and diagnose breast cancer in a timely manner. However, breast tumors in ultrasound images are characterized by irregular shapes, large-scale changes, and blurry boundaries, which bring great challenges to the segmentation and classification of breast ultrasound images. In this paper, we propose an approach for ultrasound image segmentation and classification of Triple Negative Breast Cancer (TNBC) and Non-Triple Negative Breast Cancer (NTNBC) based on deep learning and radiomics. Regarding the deep learning framework, we propose a benign and malignant segmentation model based on the U-Net model for Channel and Spatial Efficient U-Net (CSE-U-Net) breast ultrasounds. First, during the skip connection stage, to obtain more spatial and contextual semantic information, we implemented a channel attention module (CM) and spatial awareness module (SM). Then, Efficient Channel Attention (ECA) is added to the feature extraction part of the U-Net backbone network to ensure model robustness. Regarding the radiomics, we extract image features using histogram, texture, and filter features, and compare them using six machine learning classification algorithms. Experimental results indicate that CSE-U-Net is more effective and accurate in many evaluation indexes; among the six Machine Learning (ML) algorithms evaluated, the extreme gradient boosting algorithm was able to classify breast cancer more accurately into TNBC and NTNBC than the other five algorithms evaluated. The prediction results show that the CSE-U-Net and ML algorithms are efficient and can be used for the segmentation and classification of breast cancer into TNBC and NTNBC.
Autism spectrum disorder (ASD), characterized by social communication deficits and restricted/repetitive behaviors, faces limitations in conventional behavioral diagnostics. We propose BioDiag-Net, a deep learning-based dynamic graph convolutional network that integrates spatial brain topology with resting-state fMRI (rs-fMRI) functional connectivity. The architecture adaptively fuses whole-brain features with default mode network (DMN) dynamics through mutual information-driven selection, effectively suppressing noise while preserving neurobiologically relevant patterns. Evaluated on the ABIDE I dataset (1,035 multisite rs-fMRI scans), BioDiag-Net achieves 83.48% accuracy in ASD classification, capturing both global network reorganization and localized DMN aberrations. Biomarker visualization enables interpretable neuroimaging diagnosis, revealing neurophysiological substrates, providing interpretable diagnostic evidence while demonstrating transfer potential for cross-disorder neuroimaging analysis. This work bridges computational neuroscience with clinical practice, advancing personalized psychiatric diagnostics.
Pulmonary nodule segmentation plays a crucial role in the early detection of lung cancer. However, existing detection methods fail to efficiently extract multi-scale features and precisely reconstruct nodule boundaries, especially when dealing with nodules of varying sizes, irregular shapes, and complex backgrounds. To address this challenge, we introduce a novel U-Net architecture called EMCA-UNet (Efficient Multi-scale Convolution Attention). Specifically, the encoder of EMCA-UNet comprises Residual Multi-scale Attention Convolution (RESMAC) blocks, which enhance the model's feature extraction capabilities. The decoder integrates three modules: Multi-scale Convolution Attention (EMCA), Large Kernel Attention Gates (LGAG), and Efficient Up-Convolution Blocks (EUCB). These modules synergistically form a new multi-scale attention decoding layer that replaces the traditional U-Net decoder structure, thus enabling efficient multi-scale feature fusion and precise boundary reconstruction. Research has demonstrated that EMCA-UNet outperforms traditional models on publicly available datasets, such as LIDC-IDRI and LNDb. The Dice coefficient improved from 0.9198 to 0.9304, and the IoU increased from 0.8727 to 0.8856. Experimental results demonstrate that the proposed method offers a novel perspective for pulmonary nodule segmentation.
Object pose estimation is a critical task in the field of machine vision. Existing pose estimation methods often suffer from challenges such as large parameter sizes, complex architectures, and high computational costs, which limit their applicability in real-world scenarios. To address these issues, we propose a novel category-level object pose estimation model, named MSPF-LMFF. This model eliminates the reliance on attention mechanisms or precise 3D models, significantly reduces computational complexity, and enhances pose estimation accuracy, demonstrating superior performance on both real and synthetic datasets. Specifically, the MSPF module enriches the features of point clouds by integrating multi-scale image texture features with prior point cloud features, making them closer to the target object point cloud. Subsequently, the LMFF module combines geometric features of fused point cloud, depth image features, and geometric features of the target object point cloud to enhance the robustness of the model. At the same time, this module fuses adaptive point cloud features with the target object’s geometric features to improve the reliability of shape information, thereby enhancing the model’s generalization capability across different instances of the same category. Following this, a multi-layer perceptron (MLP) generates deformation and mapping matrices to reconstruct the target object’s normalized object coordinate space (NOCS) model. Finally, based on the NOCS model, the point cloud registration module computes the target object’s 6D pose and 3D dimensions. Experimental results demonstrate that MSPF-LMFF outperforms existing methods on the NOCS-REAL and NOCS-CAMERA datasets while significantly reducing parameter sizes and training time. Moreover, the proposed model exhibits exceptional generalization capabilities on the Wild 6D dataset, further validating its effectiveness.
Motivated by the escalating demand for increasingly immersive and realistic virtual reality experiences, the enhancement of depth perception via precise 3D reconstruction has emerged as a fundamental requirement. In response to existing challenges such as depth ambiguity, calibration inaccuracies, and occlusion handling, this paper introduces a novel real-time binocular 3D reconstruction framework based on the integration of feature-based and learning-based process. Specifically, the proposed approach features a hybrid neural network improving depth accuracy in textureless areas and a deep learning model incorporating local geometric priors to refine disparity maps, thereby enhancing depth prediction in occluded regions of the inputs. Experimental results demonstrate significant improvements in the accuracy and efficiency of real-time binocular 3D reconstructions, with the system capable of producing high-quality, detailed models in real time. This work not only addresses some key limitations of current technologies but also significantly advances the potential for more sophisticated and immersive VR applications.
The human heart abnormality can be detected using electrocardiogram (ECG) signal which is an electrical activity waveform with P, QRS and T waves. ECG abnormality normally analyzed using shape, wave form peaks and time duration of waves etc. Traditional techniques require manual interpretation to recognize the hear abnormality, hence time consuming process. In this research, automated deep learning model that understand each patches in ECG wave for accurate abnormality classification is implemented. We propose novel patch segmentation with a 180-timeframe timestamps for best feature engineering and learning of ECG waveforms in the MIT-BH dataset. The segmented features support the model to focus on main features, which helps to improve the learning efficiency and prediction accuracy of machine learning and deep learning algorithms. Waves are segmented as patches in 180 time stamps, and features are extracted for learning and classification. A machine learning model random forest with SMOTE and a hybrid deep learning model CNN-LSTM with SMOTE are used to train and test the patched ECG waveform. Further to test the generalizability of model, we tested the PTBXL dataset from physio net to MIT trained CNN-LSTM model. The result shows that our proposed model on random forest acquired a maximum of 98.3 % accuracy in binary classification and 99.2% in multiclass classification. CNN-LSTM with patch segmentation achieves 96.5% accuracy. For PTBXL dataset testing we achieved 91% of accuracy.
In the era of industrial automation and smart manufacturing, the reliability of industrial equipment is of paramount importance as equipment failures can trigger severe production disruptions and safety hazards. While Bayesian Learning effectively handles uncertainties and enhances fault detection interpretability, it faces challenges in data privacy and centralized processing, whereas Federated Learning safeguards data privacy. Bayesian Federated Learning (BFL) integrates their strengths for privacy-preserving probabilistic modeling with superior generalization. This paper innovatively proposes BFL-CI (Bayesian Federated Learning with Channel Importance), presenting a comprehensive framework and detection process that analyzes channel parameter distributions and transmits selected channel distributions as prior distributions for subsequent clients. Experiments on bearing and gearbox datasets simulating multi-sensor scenarios verify BFL-CI’s effectiveness in improving fault detection accuracy and robustness. This paper fills the gap in applying BFL to industrial equipment fault detection, offering a cutting-edge privacy-preserving solution for reliable intelligent industrial monitoring.
In recent years, the demand for multimodal data retrieval has been growing rapidly. As two major modalities for information transmission, images and texts exhibit significant differences in feature distribution. To address challenges in image-text retrieval-such as balancing efficiency with performance and enhancing semantic modelling-this paper proposes an efficient cross-modal feature matching model based on the CLIP framework, including two parts: feature extraction and contrastive learning. During feature extraction, pre-trained VIT and BERT models are used to capture deep semantic features of images and texts, which achieve significant improvements in Feature Entropy (text: 4.27 vs. 3.62; image: 4.13 vs. 3.47) and Mutual Information (28.3% for text, 31.5% for image) compared with the baseline, indicating stronger semantic expressiveness and alignment. Through contrastive learning with the cosine-based loss function and Adam optimization, the model ensures stable convergence. Furthermore, preprocessing innovations such as removing redundant text tokens and Base64 image encoding boost training efficiency. Experiments on a dataset of 50,000 image-text pairs demonstrate that our model achieves high and stable retrieval performance with R@1, R@5, and R@10 scores ranging from 80% to 90%. Compared to the classic DeViSE model, our approach yields improvements of 12.9%, 10.0%, and 9.0% across the three metrics, confirming the model's superior accuracy and generalization in large-scale retrieval scenarios. Finally, the model is evaluated on image-text retrieval tasks, where it consistently demonstrates strong cross-modal matching capabilities and accurately captures the semantic associations between images and texts.
This paper establishes a combined fire recognition and prediction model to study the spread of forest fires in response to the frequent occurrence of forest hill fires and its difficult recognition and prediction pain point. Based on the traditional recognition neural network, this paper innovatively establishes Fully Convolutional Network (FCN) to improve the fire recognition accuracy. The Rothermel model is introduced for fire spread prediction of the Palisades Mountain Fire in Los Angeles, and it is found that the accuracy of the Rothermel model is as high as 87% and the stability is about 70.9%. Referring to the excellent model performance of the Rothermel model, this paper establishes a combined model for identification and prediction with the combination of FCN and Rothermel in order to improve the accuracy of fire identification and prediction, and provide double accuracy to ensure the reliability of the study. Based on the simulation of forest fires in a simulated wildland environment, the fire spreading stages are segmented into 4 parts, fire recognition is performed by FCN network, and Rothermel fire prediction is performed based on the recognition results. It is found that the combined model effectively reduces the errors of individual models, complements the advantages of individual models, and improves the fire identification and prediction accuracy. At last, this paper suggests a combination with the field of drones for smart fire prevention and reference.
This paper proposes Efficient Screening-Based Optimization (ESO), a dual-threshold greedy screening framework for large-scale sparse learning. ESO integrates adaptive feature evaluation with dynamic parameter updates to address computational inefficiency in ultra-sparse scenarios. By employing a probabilistic screening mechanism and proximal-based test functions, it achieves 50-70% faster computation than state-of-the-art methods when regularization parameters approach 105. Experiments on synthetic and real-world datasets demonstrate robustness across penalty functions (L1, SCAD, MCP) and data types (image, genomic). Theoretical analysis confirms solution consistency, while parameter sensitivity studies guide practical implementation. The method significantly enhances scalability for high-dimensional problems.
