173 pubs · 2,255 cites · h-index 25
Experimental study demonstrates rapid online orthophoto generation across collaborative drone swarms, indicating scalable aerial mapping with bounded latency.
Observational analysis shows time series forecasting improved using a dual-stream transformer model, indicating enhanced adaptability.
Recent advancements in deep-learning and computer vision technologies, coupled with the availability of large-scale remote-sensing image datasets, have accelerated the progress of remote-sensing object detection. However, large-scale remote-sensing images typically feature extensive and complex backgrounds with small and sparsely distributed objects, which pose significant challenges to detection performance. To address this, we propose a novel framework for accurate object detection, termed transparent mask background optimization for accurate object detection (TMBO-AOD), which incorporates a clear focus module and an adaptive filtering framework. The clear focus module constructs an empirical background pool using a Gaussian distribution and introduces transparent masks to prepare for subsequent optimization stages. The adaptive filtering framework can be applied to anchor-based or anchor-free models. It dynamically adjusts the number of candidates generated based on background flags, thereby optimizing the label assignment process. This approach not only alleviates the imbalance between positive and negative samples but also enhances the efficiency of candidate generation. Furthermore, we introduce a novel separated loss function that strengthens both foreground and background consistencies. Specifically, it focuses the model’s attention on foreground objects while enabling it to learn the consistency of background features, thus improving its ability to distinguish objects from the background. We employ YOLOv8 combined with our proposed optimizations to evaluate our model in many datasets, demonstrating improvements in both accuracy and efficiency. Additionally, we validate the effectiveness of our adaptive filtering framework in both anchor-based and anchor-free methods. When implemented with YOLOv5 (anchor based), the framework reduces the candidate generation time by 48.36%, while the YOLOv8 (anchor-free) implementation achieves a 46.81% reduction, both with maintained detection accuracy.
In this article, we investigate the paradigm of deep learning techniques to enhance the performance of visual-based simultaneous localization and mapping (vSLAM) systems, particularly in challenging environments. By leveraging deep feature extraction and matching methods, we propose a robust, versatile hybrid visual SLAM framework, Rover-SLAM, aimed at improving adaptability in adverse conditions, such as dynamic lighting conditions, areas with weak textures, and camera jitter. Building on excellent learning-based algorithms of recent years, we designed from scratch a novel system that uses the same feature extraction and matching approaches for all SLAM tasks. Our system supports multiple modes, including monocular, stereo, monocular-inertial, and stereo-inertial configurations, offering flexibility to address diverse real-world scenarios. Through comprehensive experiments conducted on publicly available datasets and self-collected data, we demonstrate the superior performance of our Rover-SLAM system compared to the state-of-the-art (SOTA) approaches. We also conducted an in-depth analysis of the integration of visual SLAM with deep learning methods quantitatively to provide insights for future research endeavors in this domain. The experimental results showcase the system’s capability of achieving higher localization accuracy and robust tracking performance.
Periodontal disease is the most common type of oral disease. Periodontal bone defect is the clinical outcome of advanced periodontal disease, which seriously affects the quality of life of patients. Promoting periodontal tissue regeneration and repairing periodontal bone defects is the ultimate treatment goal for periodontal disease, but the means and methods are very limited. Hydrogels are a class of highly hydrophilic polymer networks, and their good biocompatibility has made them a popular research material in the field of oral medicine in recent years. This paper reviews the current mainstream types and characteristics of hydrogels, and summarizes the relevant basic research on hydrogels in promoting periodontal tissue regeneration and bone defect repair in recent years. The possible mechanisms of action and efficacy evaluation are discussed in depth, and the application prospects are also discussed.
In hyperspectral image (HSI) classification tasks, extracting surface features is crucial, but the complexity of HSI brings considerable challenges. Traditional convolutional neural networks (CNNs) and vision transformer (ViT) perform well but have large parameters and are time-consuming. In contrast, graph convolutional neural networks (GCNs) are fast to train with few parameters. The simple linear iterative clustering (SLIC) segmentation algorithm generates pixel-level graph data and superpixel-level graph data. Single-branch GCNs learn only one type of graph data resulting in insufficient feature richness. In two-branch GCNs, one branch extracts local spatial-spectral features from pixel-level data, another branch extracts global spatial-spectral features from superpixel-level data, and finally, the two branches are combined. However, the existing two-branch GCNs do not focus on the close fusion of local and global features. Therefore, in this article, we propose new two-branch GCNs with local and global feature fusion (LG-GCNs). The LG-GCN model proposed in this study not only has a short training time but also has a high accuracy. The LG-GCN can capture more comprehensively the local and global spatial-spectral semantic information of HSI. The experimental results on three publicly available HSI datasets show that LG-GCN outperforms other state-of-the-art models in terms of classification accuracy and robustness. In addition, it has a shorter training time and is more resistant to noise. The source code of the LG-GCN model is available athttps://github.com/Dreamvai/LG-GCN.
Traditional training models often experience significant performance drops on test sets when the training and testing data distributions differ. To address this domain shift problem, we propose D3Net, a domain generalization (DG) classification network for hyperspectral images (HSIs) based on generative adversarial networks (GANs). Specifically, D3Net consists mainly of a generator and a discriminator. The generator extracts domain-invariant information from the source domain to generate data with core classification features, while the discriminator employs a dual-confidence model to enhance the capture of domain-invariant features. Through adversarial iterations, the model is able to adapt to the domain shift effects of unknown data. Unlike existing DG methods that rely on random perturbations for data augmentation, D3Net utilizes learnable convolutional neural networks (CNNs) rather than randomization to enhance the model’s learning capability. We conducted cross-scene classification experiments on datasets from Houston, Pavia, and Indiana, and the results demonstrate the effectiveness of our approach. The code for D3Net is available at:https://github.com/gmsjzyq123/D3Net.
With the widespread use of GPS-enabled smart devices and the increased availability of wireless networks, mobile crowdsourcing system (MCS) has recently been proposed as a framework that automatically requests workers to perform location sensitive tasks. In the task allocation problem of MCS, existing algorithms lack consideration of the impact of nonadjacent and discontinuous execution of tasks on task allocation. Workers may have different task preferences in different time periods. Nonadjacent and discontinuous tasks provide important correlations for understanding workers’ behavior. This article introduces a novel task allocation algorithm based on spatiotemporal attention network (STATA). STATA takes into account factors such as the spatiotemporal distribution of tasks and workers as well as the location preferences and abilities of workers, and integrates them into a unified network for modeling. First, all historical tasks performed by workers are aggregated to obtain the correlation of all historical tasks. Then, the most plausible candidate tasks are recalled from the weighted representation for allocation. STATA utilizes the spatiotemporal attention mechanism to capture the relationship between these factors, ultimately improving the accuracy of task allocation. Extensive experiments demonstrate that the STATA model exhibits superior performance in terms of task allocation accuracy and practical application capabilities.
In this letter, we propose a robust cross-drone multi-target association algorithm that aims to associate the identities of targets detected by different drones with overlapping FOVs. Unlike other methods using image features (target appearance features and background features), our approach makes full use of the three-dimensional spatial consistency (TSC) between targets. It has wider applicability in deployed overlapping cross-drone systems with unreliable image features. Specifically, we first explore target correlation from the perspective of three-dimensional space, and propose an association method that integrates position and altitude consistency constraints, without relying on image features. After that, we mine topological mapping relationships (TMR) between cameras from the preliminary results and propose a TMR-based re-association approach using cost evaluation to optimize the preliminary results, which effectively improve association accuracy. Extensive experimental evaluations on Airsim-based dataset verify the effectiveness and robustness of our proposed method.
With the increasing popularity of wireless networks and the development of smart cities, the Mobile Crowdsourcing System (MCS) has emerged as a framework for automatically assigning spatiotemporal tasks to workers. The study of mobile crowdsourcing makes a valuable research contribution to community service and urban route planning. However, previous algorithms have faced challenges in effectively addressing task allocation issues with massive spatial data. In this paper, we propose a novel solution to the spatiotemporal task allocation problem using a knowledge graph. Firstly, we construct a robust spatiotemporal knowledge graph (STKG) and employ a knowledge graph embedding algorithm to learn the representations of nodes and edges. Next, we utilize these representations to build a task transition graph, which is a weighted and learning-based graph that highlights important neighbors for each task. We then apply a simplified Graph Convolutional Network (GCN) and an RNN-based model to enhance task representations and capture sequential transition patterns on the task transition graph. Furthermore, we design a similarity function to facilitate personalized task allocation. Through experimental results, we demonstrate that our solution achieves higher accuracy compared to existing approaches when tested on three real datasets. These research findings are significant as they contribute to an 18.01% improvement in spatiotemporal task allocation accuracy.
SAR-ATR (synthetic aperture radar-automatic target recognition) is a hot topic in remote sensing. This work suggests a few-shot target recognition approach (FTL) based on the concept of transfer learning to accomplish accurate target recognition of SAR images in a few-shot scenario since the classic SAR ATR method has significant data reliance. At the same time, the strategy introduces a model distillation method to improve the model’s performance further. This method is composed of three parts. First, the data engine, which uses the style conversion model and optical image data to generate image data similar to SAR style and realize cross-domain conversion, can effectively solve the problem of insufficient training data of the SAR image classification model. Second is model training, which uses SAR image data sets to pre-train the model. Here, we introduce the deep Brownian distance covariance (Deep BDC) pooling layer to optimize the image feature representation so that the model can learn the image representation by measuring the difference between the joint feature function of the embedded feature and the edge product. Third, model fine-tuning, which freezes the model structure, except the classifier, and fine-tunes it by using a small amount of novel data. The knowledge distillation approach is also introduced simultaneously to train the model repeatedly, sharpen the knowledge, and enhance model performance. According to experimental results on the MSTAR benchmark dataset, the proposed method is demonstrably better than the SOTA method in the few-shot SAR ATR issue. The recognition accuracy is about 80% in the case of 10-way 10-shot.
Click-through rate (CTR) prediction is a research point for measuring recommendation systems and calculating AD traffic. Existing studies have proved that deep learning performs very well in prediction tasks, but most of the existing studies are based on deterministic models, and there is a big gap in capturing uncertainty. Modeling uncertainty is a major challenge when using machine learning solutions to solve real-world problems in various domains. In order to quantify the uncertainty of the model and achieve accurate and reliable prediction results. This paper designs a CTR prediction framework combining feature selection and feature interaction. In this framework, a CTR prediction model based on Bayesian deep learning is proposed to quantify the uncertainty in the prediction model. On the squeeze network and DNN parallel prediction model framework, the approximate posterior parameter distribution of the model is obtained using the Monte Carlo dropout, and obtains the integrated prediction results. Epistemic and aleatoric uncertainty are defined and adopt information entropy to calculate the sum of the two kinds of uncertainties. Epistemic uncertainty could be measured by mutual information. Experimental results show that the model proposed is superior to other models in terms of prediction performance and has the ability to quantify uncertainty.
Multivariate time series prediction models perform the required operation on a specific window length of a given input. However, capturing complex and nonlinear interdependencies in each temporal window remains challenging. The typical attention mechanisms assign a weight for a variable at the same time or the features of each previous time step to capture spatio-temporal correlations. However, it fails to directly extract each time step's relevant features that affect future values to learn the spatio-temporal pattern from a global perspective. To this end, a temporal window attention-based window-dependent long short-term memory network (TWA-WDLSTM) is proposed to enhance the temporal dependencies, which exploits the encoder-decoder framework. In the encoder, we design a temporal window attention mechanism to select relevant exogenous series in a temporal window. Furthermore, we introduce a window-dependent long short-term memory network (WDLSTM) to encode the input sequences in a temporal window into a feature representation and capture very long term dependencies. In the decoder, we use WDLSTM to generate the prediction values. We applied our model to four real-world datasets in comparison to a variety of state-of-the-art models. The experimental results suggest that TWA-WDLSTM can outperform comparison models. In addition, the temporal window attention mechanism has good interpretability. We can observe which variable contributes to the future value.
Earthquake prediction based on extreme imbalanced precursor data is a challenging task for standard algorithms. Since even if an area is in an earthquake-prone zone, the proportion of days with earthquakes per year is still a minority. The general method is to generate more artificial data for the minority class that is the earthquake occurrence data. But the most popular oversampling methods generate synthetic samples along line segments that join minority class instances, which is not suitable for earthquake precursor data. In this paper, we propose a Safe Zone Synthetic Minority Oversampling Technique (SZ-SMOTE) oversampling method as an enhancement of the SMOTE data generation mechanism. SZ-SMOTE generates synthetic samples with a concentration mechanism in the hyper-sphere area around each selected minority instances. The performance of SZ-SMOTE is compared against no oversampling, SMOTE and its popular modifications adaptive synthetic sampling (ADASYN) and borderline SMOTE (B-SMOTE) on six different classifiers. The experiment results show that the quality of earthquake prediction using SZ-SMOTE as oversampling algorithm significantly outperforms that of using the other oversampling algorithms.
Multi-label learning is dedicated to learning functions so that each sample is labeled with a true label set. With the increase of data knowledge, the feature dimensionality is increasing. However, high-dimensional information may contain noisy data, making the process of multi-label learning difficult. Feature selection is a technical approach that can effectively reduce the data dimension. In the study of feature selection, the multi-objective optimization algorithm has shown an excellent global optimization performance. The Pareto relationship can handle contradictory objectives in the multi-objective problem well. Therefore, a Shapley value-fused feature selection algorithm for multi-label learning (SHAPFS-ML) is proposed. The method takes multi-label criteria as the optimization objectives and the proposed crossover and mutation operators based on Shapley value are conducive to identifying relevant, redundant and irrelevant features. The comparison of experimental results on real-world datasets reveals that SHAPFS-ML is an effective feature selection method for multi-label classification, which can reduce the classification algorithm's computational complexity and improve the classification accuracy.
Negative selection algorithm(NSA), which is the most representative classification algorithm among immune heuristic algorithms, has been successfully applied to solve many classification problems. However, there are two obvious problems with NSA. First, NSA establishes specific antibodies for each antigen based on the mechanism of specific antigen matching antibody, which leads to too many and redundant detectors. Second, the detection stage needs to calculate the matching degree of the antigen with all the detectors, and the detection efficiency is low. This paper proposes a new non-specific natural killer cell algorithm (NKA) based on pathogen dose. NKA draws on the mechanism of NK cells constructing phenotype detectors based on pathogen dose. First, NKA defines dose and phenotype detector, and optimizes it based on the memory evolution mechanism of phenotype; then establishes k-d tree for the optimized phenotype, and pathogens only need to match the dose with the nearest phenotype detector. Experimental results show that the method proposed in this paper, NKA, can not only achieve a better performance through fewer detectors, but also has a higher efficiency in the training and detection phase, compared with three NSA-based algorithms.
Online task assignment is one of the core research issues of spatio-temporal crowdsourcing technology. The current researches on minimizing travel cost all focus on the scenario of two objectives (task requesters and workers). This paper proposes a two-stage framework (GH) based on Greedy algorithm and Hungarian algorithm for three-objective online task assignment to minimize travel cost. In order to further optimize the efficiency and average travel cost, this paper proposes GH-AT (Adaptive Threshold) algorithm based on GH algorithm, and redesigns the Hungarian algorithm into the sHungarian algorithm. sHungarian algorithm has lower time complexity than Hungarian algorithm. sHungarian algorithm is not only suitable for the problem studied in this paper, but also for all task assignment problems with constraints. Compared with Greedy algorithm, GH-AT algorithm has lower travel cost and higher total utility. In terms of the number of matches, GH-AT is slightly lower than Greedy algorithm. In terms of time cost, GH-AT algorithm is higher than Greedy algorithm, but much lower than GH algorithm.
The Internet of Vehicles (IoV) is the extension of the Internet of Things (IoT) technology in the field of transportation systems. Ride‐sharing is one of intelligent travel applications in IoV. Ride‐sharing is aimed at taking passengers with similar itineraries and time arrangements to travel in the same car according to a certain matching rule. To effectively integrate transport capacity resources and reduce the number of cars on the road, ride‐sharing has become a popular and economical way of travel. The matching and optimizing of drivers and passengers are the core contents of a ride‐sharing application system. This paper mainly studies the dynamic real‐time matching of passengers and drivers in IoV, considering the main factors such as travel cost, car capacity, and utility. The matching problem is formulated in a ride‐sharing system as a Role‐Based Collaboration (RBC). A new utility method for the matching optimization of ride‐sharing is present. In this paper, we establish a model for simulating the assignment of ride‐sharing with the help of the Environments‐Classes, Agents, Roles, Groups, and Objects (E‐CARGO) model. The objective function and formal definitions are proposed. The utility and time of optimal matching are obtained by using the Kuhn‐Munkres algorithm on the revenue matrix. The experimental results show that the proposed formal method based on the E‐CARGO model and utility theory can be applied in the ride‐sharing problem. Numerical experiments show that the matching time cost increases with the increase of the number of drivers and passengers participating in the ride‐sharing system. When the number of drivers and passengers is different, one‐to‐many matching takes the least time, and one‐to‐one matching takes more time. When the number of drivers and passengers is the same, the time cost of one‐to‐one matching increases sharply with a certain value (bigger than 800). Compared with other matching methods, the time spent by the one‐to‐many method is reduced by 30%. The results show that the proposed solution can be applied to the matching and pricing in a ride‐sharing system.
In the whole online auction industry, it is important to predict end prices. To improve the accuracy of predicting end prices using less training data, this paper proposes a hybrid algorithm which combines multiple linear regression with Kalman filter (MLRKF). The proposed algorithm solves problems of low prediction accuracy and over fitting when we signally use multiple linear regression algorithm to predict in machine learning. Firstly, multiple linear regression and Kalman filter models are introduced and analyzed. Secondly, we view the prediction problem with multiple linear regression as a weight parameter optimization problem, and demonstrate the method in theory. Then MLRKF prediction model is provided. Finally, the proposed model is used to predict eBay end prices based on two datasets. In our experiment, MLRKF prediction model has been compared with other models including multiple linear regression, multiple linear ridge regression, Lasso, random forest, support vector machine, and recurrent neural network. This hybrid algorithm has been proved to produce highly accurate results with less training data and lower time cost. The experimental results indicate that the proposed algorithm has a small error rate by calibration metrics in behavioral bidding tasks.
Click-through rate (CTR) prediction is an important issue in online advertising and recommendation systems. It is used to estimate the likelihood that a user will click on ads. The method used in the traditional CTR prediction task is to improve the prediction effect through a large number of feature engineering. However, these methods are time-consuming and laborious, and the construction process is not universal. Due to the sparse and high-dimensional characteristics of data features, it is necessary to measure the importance of sparse features and obtain efficient feature interactions. In this paper, a novel CTR model based on Mutual Information and Feature Interaction (MiFiNN) is proposed. First, the mutual information of each sparse feature and the click result is calculated as the weight of each sparse feature. Subsequently, an interactive method combining the outer product and inner product is constructed to carry out the feature interaction. Then, the resulting feature interactions and the dense features set of the original input are taken as DNN inputs. We verify the proposed model on four datasets. In addition, several widely known models are introduced for comparison. The experimental results indicate the superiority of the model.