Search

In few-shot learning, the Brownian distance covariance improves classification accuracy by enhancing feature embeddings, but it does not focus on the issue of sample-related features in classification. This paper proposed the deep masked Brownian distance covariance method that generated query-guided masks based on high-dimensional semantic relationships between each pair of query and support samples, and employed the masked Brownian distance covariance matrix as the image features. Under the 5way-1shot and 5way-5shot scenarios, it carried out validation and evaluation on CUB-200-211, MiniImageNet, and Tiered-ImageNet datasets. The experiments show that this method achieves superior classification accuracy. [ABSTRACT FROM AUTHOR]

This paper proposed a watermarking algorithm for neural radiance fields (NeRF) to address the copyright protection issue of implicit 3D models. The algorithm embedded watermarks into the images in the training set through an embedding network, and then utilized the NeRF model for 3D model modeling. It generated a backdoor image by the copyright verifier from a new perspective by providing a secret perspective as input to the neural radiation field. Subsequently, it designed a watermark extractor using the hyperparameterization method of the neural network to obtain the embedded watermark image from that perspective. In a black box scenario, once there was suspicion of unauthorized use of the 3D model, the verifier could extract watermarks from a secret perspective to verify network copyright. The experimental results demonstrate that the proposed algorithm not only yields good visual effects when extracting watermarks, but also exhibits robust resistance to different types of noise attacks. [ABSTRACT FROM AUTHOR]

The scarcity of 3D multimodal data results in a lack of semantic consistency between text and visual features during supervised training using traditional methods. Meanwhile, traditional methods also overlook local relationships and global information, resulting in poor performance. To address the above issues, this paper proposed a semantic consistency constrain and local-global aware multi-modal 3D visual grounding method. Firstly, the method helped the 3D model extract point cloud-text semantic consistency features by distilling 2D pre-trained visual language model knowledge. Secondly, it designed a localglobal aware module to continuously supplement and enhanced candidate target features to match targets more accurately. Experiments conducted on the ScanRefer dataset show that the proposed method achieves 50.53% and 37.67% in terms of Acc@ 0.25 IoU and Acc@0.5 IoU and exceeds most existing 3D visual grounding methods, confirming the effectiveness of the method. [ABSTRACT FROM AUTHOR]

When using location-based services, users need to provide their real location information, which may lead to the leakage of user information. Most research only focuses on the privacy protection of individual users, while ignoring the correlation among multiple users. This paper proposed a differential privacy trajectory protection scheme based on user correlation for trajectory privacy protection issues involving multiple users. Firstly, it constructed a historical trajectory tree and used a variable-order Markov model to predict user trajectories, generating a set of highly usable trajectory datasets from the collection of trajectories. Secondly, it obtained a set of predicted trajectories with lower correlation based on the inter-user trajectory correlations. Finally, by customizing the privacy budget method, it dynamically adjusted the privacy budget for each location point according to different user privacy needs and added Laplacian noise to the published trajectories. Experimental results show that compared to the LPADP algorithm, this algorithm improves execution efficiency by 10%~15.9%. Compared to both PTPP and LPADP algorithms, it enhances data usability by 11%~16.1%, while also increasing the level of privacy protection. [ABSTRACT FROM AUTHOR]

MSs-MEC can adaptively adjust service cache decisions according to needs, so that edge servers deploys on the user side can flexibly handle tasks of different service types. However, in practical applications, the successful migration of specific types of tasks depends on the early installation of the service environment. In addition, simultaneous task migration and service caching may result in time conflicts and computation delays. Therefore, in response to the above issues, this paper firstly decoupled task migration and service caching decisions. To address the shortcomings of DRL in high-dimensional mixed decision spaces, where performance improvement was not significant(such as low utilization rate during resource allocation), it combined DRL with Transformer to learn from historical data and output task migration decisions for the current time slot and the next time slot, ensuring that tasks could be executed immediately when they reached the edge server. Secondly, in order to improve resource utilization in resource allocation problems, it decomposed the problem into continuous resource allocation problems and discreted task migration and service caching problems, and used convex optimization technique to solve the optimal decision for resource allocation. Extensive numerical results indicate that the proposed algorithm can effectively reduce the average completion delay of tasks compared to other baseline algorithms, at the same time, it also has excellent performance in resource utilization and stability. [ABSTRACT FROM AUTHOR]

Tandem repeat sequences are difficult part in genome construction, due to the high similarity between repeated units and the ambiguity in copy numbers, it often result in multiple candidate positions during sequence alignment. The challenge lies in rapidly and accurately filtering out the correct alignment positions. Existing methods address this issue by using seeds (short sequences selected from sequencing fragments) to locate and extend candidate alignment positions, but overlook the distinctive characteristics of tandem repeat sequences when selecting seeds. To tackle the problem, this paper proposed a position filtering method for tandem repeat sequence alignment, which filtered alignment results by calculating the similarity of rare kmer sequences. Additionally, it implemented a strategy of merging rare kmers to expedite computation, coupled with a fuzzy search based on edit distance to enhance filtering information density. Experimental results demonstrate that this approach improves both the recall and accuracy of alignment results on simulated datasets while achieving approximately a 2-fold increase in computational speed compared to existing methods, with notable parallel acceleration effects. [ABSTRACT FROM AUTHOR]

The existing legal judgment prediction methods based on knowledge graph focus on the element entities and relationships of the case, and cannot adequately capture the characteristic information of the case. Aiming at this problem, the paper proposed a knowledge graph legal judgment prediction method that enhanced the fusion of case features. Firstly, this method used bidirectional gated recurrent neural network to mine the deep semantic feature information such as causality and time sequence of fact description text. Then, it calculated the feature representation of the learning class case by the similarity attention between cases in the knowledge graph vector space. Finally, the fusion of feature information and structured knowledge of knowledge graph enriched the semantic feature representation of entities and relationships in the case fact text, and realized the legal judgment link prediction task. The experimental results on the two types of crime datasets of dangerous driving and theft show that the method improves the two key evaluation indicators of MRR and Hit@1 by about 1.5% compared with the current best-performing link prediction models. The indicators such as Hit@3 and Hit@ 10 are also improved, which verifies that the case feature enhancement fusion can supplement the missing case feature information in the legal knowledge graph and improve the prediction effect. [ABSTRACT FROM AUTHOR]

Because the accuracy of log parsing is not high and the lack of tag samples reduces the accuracy of anomaly detection, this paper proposed a new semi-supervised anomaly detection method based on logs. Firstly, the method enhanced the log parsing method of the dictionary to retain parameter information in log events, improving the utilization and accuracy of log resolution. Next, the method utilized BERT to encode semantic information in the template, obtaining the semantic vector of the log. Then, the method employed the clustering method to estimate the tag, which effectively alleviated the problem of insufficient data labeling and enhanced the model's ability of detecting unstable data. Finally, the method captured context information from two directions based on the bidirectional temporal convolution network (Bi-TCN) with residual blocks, which enhanced the accuracy and efficiency of anomaly detection. To evaluate the method s performance, it conducted extensive experiments on two datasets. The results demonstrate that the proposed method achieves an average improvement of 7%, 14.1% and 8.04% in F₁ value compared to the latest three benchmark models, LogBERT, PLELog and LogEncoder, enabling efficient and accurate log parsing and log anomaly detection. [ABSTRACT FROM AUTHOR]

Code clone detection is an important technology to improve software development efficiency, quality, and reliability. Single-language clone detection based on AST has achieved significant performance. However, the existence of synonyms and near-synonyms in AST nodes of cross-language codes and the high cost of manual labeling limit the effectiveness and usefulness of existing clone detection methods. To address these issues, this paper proposed a cross-language code clone detection method based on contrastive tree convolutional neural network (CTCNN). Firstly, it parsed the codes of different programming languages into ASTs, and processed the node types and values of ASTs by synonym conversion to reduce the differences between ASTs in different programming languages. At the same time, it employed contrastive learning to augment negative samples and train the model, so that this approach ensured the minimization of distances between clone pairs and the maximization of distances between non-clone pairs in small sample datasets. Finally, it evaluated the proposed method on a public dataset with precision, recall, and F1-scores of 95.6%, 99.98%, and 97.56%. The results show that compared to the best existing methods CLCDSA and C4, the proposed model improves the detection accuracy by 43.92% and 3.73%, and increases the F1score by 29.84% and 6.29%, which confirms that the proposed model is an effective cross-language code clone detection method. [ABSTRACT FROM AUTHOR]

In the disaster scene, the timely and reliable collection of first-hand and important information in the disaster area is the key to the early disaster warning research and rescue work. Unmanned aerial vehicle(UAV) is an efficient auxiliary tool for establishing emergency communication network within disaster zones. Through the investigation of the deployment methods of UAV in emergency scenarios in the existing research, this paper reported the problem that node energy efficiency was not considered in UAV deployment. Since the ground sensor nodes were located inside the disaster area in a hostile and extremely passive environment, the combination of disaster scenarios. Taking improving the energy efficiency of ground nodes as the optimization goal for the first time, based on the deep reinforcement learning method and on the basis of DDQN network model, the adaptive deployment problem of UAV for information collection in disaster scenarios was solved by defining experience playback priority, reasonably designing reward function and adopting complete decentralized training method. Simulation results show that the energy efficiency of the nodes under the proposed algorithm is 21% higher than that of the DDQN benchmark algorithm, and the training speed is 42% and 34% higher than that of the DDPG and A3C algorithms respectively. [ABSTRACT FROM AUTHOR]

Achieving unsupported 3D printing still faces challenges in handling large-angle overhangs and improving printing precision. This paper presented a cone slicing algorithm for multi-freedom unsupported 3D printing of NURBS volume parametric models. Though the geometric mapping of the 3D model file of the model, it applied a horizontal slicing algorithm to this mapped model, resulting in a continuous print path G-code of the deformed model. This G-code was then subjected to inverse mapping, generating multi-degree-of-freedom continuous printing paths suitable for six-axis robotic arms. Utilizing the conical slicing algorithm to plan printing paths enabled printing of overhanging structures without the need for support. It validates the accuracy, feasibility, and effectiveness of the conical slicing algorithm through simulation experiments and comparative trials. The introduction of the conical slicing algorithm offers a more efficient, economical, and green solution for unsupported 3D printing of NURBS-based parametric models. [ABSTRACT FROM AUTHOR]

Stock prediction is typically a non-linear time series task. However, few researchers attempt to systematically reveal the underlying structure of the stock market through technical data. The interactions of collaboration or conflicts among various business domains can explain the fluctuations in stock. The incorporation of this additional information aids in predicting the future trends of stocks. In order to represent the trading situation of the stock market as realistically as possible and to express the explicit or implicit relationships between stocks, this paper proposed a stock price prediction model sDHN based on a dynamic heterogeneous network, which synthesized the base of the stock and the industry and geographical information, and modeled it as a dynamic heterogeneous network. The model introduced dynamic time series capabilities to the network, and the algorithm creatively combined four different technical levels of similarity graphs of stock nodes to generate a rich information heterogeneous graph. Finally, it aggregated the semantic information hidden in different meta-paths to generate embeddings, exploring the potential correlations among stocks from the perspective of the heterogeneous graph. In addition, experiments on three real-world stock data sets show that the proposed model achieves accuracy improvements of between 5% and 34% over the overall baseline models. The F1-score is higher by approximately 11.5% ~37%. It demonstrates through graphical analysis the effectiveness of this approach. [ABSTRACT FROM AUTHOR]

To overcome the weaknesses of wild horse optimization algorithm, such as easy to fall into local optima and slow convergence speed, this paper proposed an enhanced wild horse optimization algorithm. Firstly, in the population initialization stage, it utilized Sinusoidal mapping to increase the diversity of the population. Secondly, in the stage update process, it designed adaptive weights with stronger nonlinear convergence to adjust the abilities of global search and local optimization. Then, it introduced perturbation factors in the leader position update stage to balance local and global exploration capabilities. Furthermore, it utilized adaptive t-distribution mutation to perturb the individual positions and improve the algorithm's ability to jump out of local optima. The effectiveness and robustness of the algorithm were validated by optimization comparisons in the CEC2021 competition test set, and the efficacy of the algorithm was verified through Wilcoxon rank sum test and MAE ranking. Finally, it applied the algorithm to two engineering problems, which verified the applicability and superiority of the algorithm for engineering optimization problems. The experimental results indicate that the enhanced wild horse optimization algorithm exhibits stronger optimization capabilities and faster convergence speed, which compared to other intelligent algorithms. [ABSTRACT FROM AUTHOR]

The vehicle traffic scheduling problem at unsignalized intersections is the research focus in the field of intelligent transportation. Because the solution space of vehicle traffic order decision problem increases exponentially with the increase of the number of vehicles, finding the optimal traffic sequence while ensuring the real-time performance has become a major problem for traffic scheduling at unsignalized intersections. To solve this problem, this paper proposed a vehicle traffic scheduling method at unsignalized intersections based on adaptive Monte Carlo tree search algorithm, which used a hierarchical framework, upper level centralized sequential decision-making, and lower level distributed trajectory planning. Firstly, the intersection model based on the conflict point was established, the connected vehicles were added to the queue to be searched, the Monte Carlo tree search process of the traffic sequence was designed according to the vehicle traffic characteristics in the intersection, the evaluation function was established with the total traffic time as the index, and then the adaptive exploration coefficient and other super parameters were designed for different traffic situations, so that the algorithm could maintain the best state in solving different vehicle numbers and different search periods. In the trajectory planning process, the acceleration two norm was taken as the objective function, and the speed, acceleration and the position of the starting point were taken as constraints to establish the optimal control proposition to solve the vehicle trajectory. Finally, experiments were carried out, and the results show that compared with other algorithms, the maximum optimization amplitude of this algorithm in numerical simulation and micro platform experiment is 33.42% and 38.04% respectively, which provides an effective solution for vehicle traffic scheduling at unsignalized intersections. [ABSTRACT FROM AUTHOR]

Aiming at two problems of existing multi-label feature selection methods: first, ignoring the influence of noise information in the process of learning label correlations; second, neglecting to explore the comprehensive label information of each cluster, the paper proposed a multi-label feature selection method that enhanced label correlation learning. Initially, it clustered the samples and treated each cluster center as a representative instance of the comprehensive semantic information of the samples, while computing its corresponding label vectors which reflected the importance of different labels contained in each cluster. Then, through the label-level self-representation of the original samples and the center of each cluster, it both captured the label correlations in the original label space, and explored the label correlations within each cluster. Finally, the self-representation coefficient matrix was sparse to reduce the effect of noise, and the original sample and the representative instance of each cluster were mapped from the feature space to the reconstructed label space for feature selection. Experimental results on nine multi-labeled datasets show that the proposed algorithm has better performance compared with other methods. [ABSTRACT FROM AUTHOR]

Federated learning as an emerging distributed computing paradigm with privacy protection, safeguards user privacy and data security to a certain extent. However, in federated learning systems, the frequent exchange of model parameters between clients and servers results in significant communication overhead. In bandwidth-limited wireless communication scenarios, this has become the primary bottleneck restricting the development of federated learning. To solve this problem, this paper proposed a dynamic sparse compression algorithm based on Z-Score. By utilizing Z-Score, it performed outlier detection on local model updates, considering significant update values as outliers and subsequently selecting them. Without complex sorting algorithms or prior knowledge of the original model updates, it achieved model update sparsification. At the same time, with the increase of communication rounds, the sparse rate was dynamically adjusted according to the loss value of the global model to minimize the total traffic while ensuring the accuracy of the model. Experiments show that in the I. I. D. data scenario, the proposed algorithm can reduce communication traffic by 95% compared with the federated average algorithm, and the accuracy loss is only 1.6%. Additionally, compared with the FTTQ algorithm, the proposed algorithm can also reduce communication traffic by 40%~50%, with only 1.29% decrease in accuracy. It proves that the method can significantly reduce the communication cost while ensuring the performance of the model. [ABSTRACT FROM AUTHOR]

Aiming at overcoming drawbacks of imbalance between the global and local search ability of a single heuristic algorithm, this paper proposed a slime mould and flower pollination hybrid algorithm based on dynamic dual population, named HASMFP. Firstly, HASMFP adopted a grouping mechanism that took the distance between individual inside population and the current optimal individual into consideration to dynamically divide whole population into slime mold subpopulation and pollen subpopulation to balance the exploration and development capabilities of the algorithm more effectively. Secondly, HASMFP used a ranking mechanism based on similarity and fitness to improve diversity of slime mold population, and further increased the probability to jump out of local optimal. Finally, HASMFP also adopted a dynamic weights and constant shrinkage coefficients with an elite guidance terms to further enhance the local and global search ability of standard flower pollination algorithm at the same time. It used 12 test functions from CEC2017 test suit as the testbed to evaluate the performance of HASMFP with other 5 algorithms: ISMA, DTSMA, HLFPA, SCFPA, and tMFPA. It conducted ablation experiments to evaluate the effectiveness of all improvement strategies applied in HASMFP. Experimental result shows that HASMFP can rank first under the combination of all improvement strategies. The result of Friedman test based on experimental data illustrates that HASMFP can achieve the supreme performance among all evaluated algorithms. [ABSTRACT FROM AUTHOR]

The current OpenIE methods cannot take into account the compactness of the extraction results and the performance of the model at the same time, which makes the extraction results unable to be better applied to downstream tasks. Therefore, this paper proposed a model that used biaffine attention for table filling and iterative extraction. Firstly, the model learned the directional information between words through biaffine attention, captured the interaction between word pairs, and then filled the two-dimensional table to make the components in the sentence share each other and identify compact components. Secondly, it used the multi-head attention mechanism to apply the representation of predicates and parameters to the context embedding, making the extraction of predicates and parameters dependent on each other and better linking the relationship components and parameter components. Finally, for sentences containing multiple relational components, it used iterative extraction to capture the inherent dependencies between each extraction without recoding. Experiments on the public datasets CaRB and Wire57 show that this method achieves higher precision and recall than baseline methods, improving F₁ values by at least 1.4% and 3.2%, while producing shorter and semantically richer extractions. [ABSTRACT FROM AUTHOR]

Addressing the challenge of suboptimal performance in existing sequential recommendation models due to severe data sparsity, this paper proposed a generative adversarial network recommendation algorithm based on reverse extension enhancement. The approach extended and enhanced interaction sequences to obtain high-quality training data, mitigating the issue of insufficient model training caused by data sparsity. Firstly, it extended the project sequences backwardly using pseudoprior terms to deepen the features of the project sequences. Secondly, it shifted the target of extension enhancement from short sequences to all user sequences, thoroughly exploring contextual information embedded in long sequences and alleviating noise issues arising from an excessively large proportion of pseudo-prior terms in augmented sequences. Finally, it employed a generative adversa-rial network with shared project embeddings, and jointly trained the discriminator and generator to enhance the model's recommendation performance. Experimental results on three public datasets demonstrate an average improvement of 30% in hit rate(HR@N) and normalized discounted cumulative gain (NDCG@N) compared to the optimal baseline ELECRec, confirming the effectiveness of reverse extension enhancement in mining sequence features and alleviating data sparsity. [ABSTRACT FROM AUTHOR]

The acoustic input network based on the Conformer encoder has the problem of insufficient extraction of FBank speech information and missing channel feature information. This paper proposed an end-to-end method based on RepVGG-SE-Conformer for speech recognition to solve these problems. Firstly, the proposed model used the multi-branch structure of RepVGG to enhance the speech information extraction capability, and using the structural re-parameterization fused the multibranch into a single branch to reduce the computational complexity and speed up the model inference. Then, based on the squeeze-and-excitation network, the channel attention mechanism made up for the missing channel feature information to improve speech recognition accuracy. Finally, the experimental results on the public dataset Aishell-1 show that the proposed method's character error rate is reduced by 10.67% compared with Conformer, and the advancement of the method is verified. In addition, the proposed RepVGG-SE acoustic input network has good generalization ability in the end-to-end scene, which can effectively improve the overall performance of speech recognition models based on Transformer variants. [ABSTRACT FROM AUTHOR]

Addressing the challenges faced by current cold start recommendation models in effectively mining structural and semantic information in heterogeneous information networks, and their tendency to overlook user behavior attributes within these networks, this paper introduced a meta-learning-based multi-view contrast fusion cold start recommendation algorithm (MVC-ML). This algorithm effectively tackled the cold start problem at both the model and data layers. Within the MVC-ML framework, it firstly extracted higher-order semantic information from heterogeneous information networks using a meta-path view. Subsequently, it captured the network's structural features using a network pattern view. Following this, the algorithm analyzed user behavior attribute information through a clustering view. Finally, MVC-ML employed a contrast learning method to integrate the information extracted from these three views, thus generating accurate representation vectors. Experimental validations on datasets, including DBook, demonstrate that the MVC-ML model, in a cold start scenario, reduces MAE by 1.67%, lowers RMSE by 2.06%, and increases nDCG@ K by 1.48% compared to traditional heterogeneous information network models such as MetaHIN. These results fully confirm the rationality and effectiveness of the MVC-ML algorithm. [ABSTRACT FROM AUTHOR]

Infrequent behavior recognition is one of the methods to reveal important information about business processes and optimize process models. Existing process discovery methods have overlooked the impact of data influence chains on infrequent behavior, resulting in some infrequent behavior being considered as noise and filtered out directly. To address this issue, this paper proposed a novel infrequent behavior analysis method based on activity recovery sets. Firstly, it filtered the event logs based on the importance of behavior and constructed an initial process model. Secondly, it extracted input and output data items of activities from transaction logs, and constructed an activity influence chain graph based on these data items. It obtained activity recovery sets based on these graphs. Finally, it calculated the behavior tolerance of each trace using the activity recovery sets to distinguish effective infrequent behavior from noise. The experimental results indicate that, compared to other methods, this study effectively distinguishes valid infrequent behaviors from noise and improves the quality of the process model in terms of fitness, precision, and simplicity. This method considers the biases caused by the activity recovery set and successfully identifies valid infrequent behaviors in event logs, thereby optimizing the process model. [ABSTRACT FROM AUTHOR]

Addressing problems such as low success rates, instability, and inefficient traffic flow in autonomous decision-making models at intersections, this paper proposed an intersection decision model based on the GA-TD3 (GRU attention twin delayed deep deterministic policy gradient) algorithm by improving TD3 algorithm from two aspects. Firstly, it introduced a memory module which using GRU neural network to improve the success rate of the decision model. Secondly, it introduced a social attention mechanism in the state space to focus on interactions with social vehicles, which ensured the stability of the model while improving the traffic efficiency of vehicles. After 20 000 rounds of training in the CARLA simulator, the TD3 algorithm achieves a success rate of 92.4%, while the success rate of the GA-TD3 algorithm is 97.6%. Additionally, the vehicle's travel time is shortened by 3.36 s. GA-TD3 algorithm improves both learning efficiency and traffic efficiency, which can alleviate traffic pressure in urban scenes and improve driving efficiency. [ABSTRACT FROM AUTHOR]

Protein complexes play a crucial role in understanding life's molecular activity process. Aiming at the problems of high false-positive/negative rate, low accuracy, and decrease in population diversity when detecting protein complexes by swarm intelligence algorithms, this paper proposed the RLDLFA-DPC. It introduced the idea of reinforcement learning to offer an adaptive level partition strategy that dynamically adjusted the level structure, solving the issue of declining population diversity in the late iteration. The algorithm also incorporated a level learning strategy where individuals learn from two excellent levels to avoid falling into a local optimum. Additionally, it utilized a local search strategy with an adaptive search radius in combination with individual and environmental information to improve the accuracy of protein complex detection. Finally, the effectiveness of the algorithm was verified by comparing it with eight classical protein complex detection methods on four datasets of saccharomyces cerevisiae proteins. [ABSTRACT FROM AUTHOR]

Understanding the algorithms described in program code can help programmers grasp the program and carry out various software engineering tasks. Because manual understanding of program algorithms requires programmers to have rich professional knowledge and is time-consuming and labor-intensive, the main research goal of program algorithm recognition is to achieve automation of program algorithm understanding, which is more efficient and accurate compared to manual understanding. This paper systematically organized the relevant work in the field of program algorithm recognition. Firstly, it combed the concepts related to program algorithm recognition and introduced methods based on knowledge representation and information retrieval. Secondly, it divided the code representation based methods into sequence based, tree based and graph based methods for detailed elaboration, and made a summary and comparison of the three types of methods. Finally, it introduced the relevant application fields of the program algorithm recognition task, analyzed the remaining problems and made prospects for future development. [ABSTRACT FROM AUTHOR]

Image stitching is an important branch in computer vision and computer graphics, and has a wide range of applications in 3D imaging and other aspects. Compared with the traditional image stitching framework based on feature point detection, the image stitching framework based on deep learning has stronger scene generalization performance. Although there are many research results on image stitching based on deep learning, there is still a lack of comprehensive analysis and summary of the corresponding research. In order to facilitate the subsequent work in this field, this paper sorted out the representative results in this field in the past 10 years. Based on the comparison between traditional stitching methods and deep learning-based image stitching methods, it collated and analysed the learning strategy and model architecture design, classical model review, and dataset from the three sub-problems of homography estimation, image stitching, and image rectangling in the research field of image stitching. It summarized some features of deep learning-based image stitching research methods and summarized the current research status in the field, and prospected the future research prospects. [ABSTRACT FROM AUTHOR]

Quantum computing is a new computing mechanism that follows the laws of quantum mechanics to control information units for calculations. It is believed to be able to achieve parallel processing capabilities on specific algorithms and has a great significance for the technological revolution in the information field. This paper summarized the progress of quantum computing applications in the financial field in recent years. Firstly, from the basic principles, it analyzed basic theories of qubits, quantum logic gates, quantum circuits, with a focus on analyzing the advantages and disadvantages of several technical routes for quantum hardware. Secondly, it summarized the research progress of quantum algorithms in derivative pricing, portfolio optimization, risk measurement, fraud detection, and market prediction, as well as the main components and architectures of quantum finance software. Finally, it discussed the talent, efficiency, and compliance challenges in the application and development of quantum computing in the financial field, as well as the prospects for future development trends. [ABSTRACT FROM AUTHOR]

PGO is a high-dimensional non-convex optimization algorithm commonly used in the back-end optimization of SLAM, which is usually modeled as maximum likelihood estimation. Since the current PGO algorithm faces difficulties in improving speed while ensuring accuracy when optimizing large-scale noise datasets, this paper proposed a hierarchical pose graph optimization algorithm based on nested dissection algorithm for large noise datasets. The algorithm firstly established a χ² test model with different distance measures, and then removed outlier points. Secondly, it used nested dissection method to split the original pose graph into a set of subgraphs and extracted a skeleton graph from these subgraphs. The skeleton represented the abstract topology of the original SLAM problem. Then, the algorithm optimized the skeleton graph and completed the initialization. Finally, experimental evaluation on simulated and real pose graph datasets show that the proposed algorithm can improve the calculation speed and scalability of the algorithm without affecting the accuracy. [ABSTRACT FROM AUTHOR]

The success of emotional dialogue systems relies on the ability to comprehend, perceive, and express emotions, while facial expressions and personality can also help. However, despite the crucial importance of this multi-modal information in multi-turn emotional dialogues, existing systems still need to be improved to leverage multi-modal information's advantages and overlook the significance of contextual relevance. To address this issue, this paper proposed a multi-turn emotional dialogue generation model based on a dua-decoding method(MEDG-DD). The model utilized a heterogeneous graph neural network encoder to integrate historical dialogue, facial expressions, emotion flow, and speaker information, obtaining a more comprehensive dialogue context. Subsequently, it employed a dual-decoding mechanism based on attention to generate emotionally rich expressions relevant to the dialogue context. Experimental results demonstrate that the proposed model effectively integrates multi-modal information, achieving more accurate, natural, and coherent emotional expressions. Compared to the traditional ReCoSa model, this model exhibits significant improvements across various evaluation metrics. [ABSTRACT FROM AUTHOR]

The graph clustering algorithm for hard samples mining is a recent research hotspot. In the current algorithm, the main problems include the lack of a fusion mechanism for comparing methods and a sample pair weighting strategy; the algorithms ignore "false negative" samples within the view when sampling positive samples and disregarding the help of graph-level information for clustering. To address the issues above, this paper proposed a graph clustering algorithm based on hard sample sampling joint contrast augmentation. Initially, it utilized an autoencoder to learn embeddings, designed a self-weighted contrast loss for representation learning by utilizing the calculated pseudo-label, similarity, and confidence information, and unified the strategies of node comparison and hard sample pair weighting across different views. By adjusting the weights of sample pairs in different confidence regions, the loss function derived the model to focus on different types of hard samples to learn discriminative features, improving the consistency of intra-cluster representation and the distinctiveness of inter-cluster representation and enhancing the ability to discriminate samples. Additionally, the clustering network projected the graph-level representation to maximize the representation consistency of clusters under different views through cluster contrast loss. Finally, combining the two comparison losses, the self-supervised training is used for iterative optimization to complete clustering. In the comparison with 9 benchmark algorithms on 5 real datasets, this algorithm achieves superior performance on 4 authoritative indicators, highlighting its excellent clustering capabilities. Ablation experiments demonstrate the effectiveness and transferability of the two contrasting modules. [ABSTRACT FROM AUTHOR]

Predictive process monitoring (PPM) relies on predictive effectiveness, and to address the challenge of improving predictive performance in PPM, this paper proposed a novel approach called behavior profile matrix enhanced business process outcome prediction. Initially, this approach extracted the behavior profile matrix by analyzing the interactions among activities and incorporated it into the model along with event sequences. Then, it used convolutional neural networks (CNN) and long short-term memory networks (LSTM) to independently capture image features from the matrix and sequence features. Finally, this approach integrated an attention mechanism to seamlessly combine both image and sequence features for predictive purpo- ses. Validation using real event logs demonstrates that the proposed enhancement method significantly enhances predictive per- formance compared to the baseline LSTM prediction methods when forecasting event log outcomes, confirming the feasibility of this approach. This approach combines the behavior profile matrix to enhance the predictive model s understanding of relation- ships between behaviors in event logs, consequently leading to an improvement in predictive performance. [ABSTRACT FROM AUTHOR]

Target detection in adverse conditions such as low illumination has always been a challenging. The factors of low light and fog can lead to reduced visibility and increased noise in images, significantly disrupting the precision of object detection. To address these issues, this paper proposed and integrated a low-light image enhancement network for machine vision perception, MVP-Net, with the YOLOv3 object detection network to construct an end-to-end enhancement detection framework, MVPYOLO. MVP-Net employed inverse mapping network technology to transform conventional RGB images into pseudo-RAW image feature space and introduced a pseudo-ISP enhancement network, DOISP, for image enhancement. The objective of MVP-Net is to harness the potential advantages of RAW images in object detection while overcoming the limitations encountered in their direct application. The model has outperformed previous works on multiple real-world low-light datasets and is adaptable to detectors with various architectures. Its end-to-end detection framework achieves a mAP(50%) metric of 78.3%, an improvement of 1.85% over the YOLO detectors. [ABSTRACT FROM AUTHOR]

Deep neural networks tend to overfit to biased training data when there are noisy labels or imbalanced class distributions in the training set. Using reweighting strategies with appropriate sample weighting is a common method to address this issue. However, improper reweighting schemes will introduce additional overhead and bias to the network's learning process, it is difficult to solve overfitting problems in biased distribution networks using only reweighting methods. To address this problem, this paper proposed a method that combined label smoothing regularization, class margin regularization, and reweighting, and presented an EMN method based on adaptive reweighting and regularization, which consisted of a base network for classification and an ensemble meta-net for hyperparameter estimation. The method first obtained the sample loss through the base network, then used three meta-learners to estimate the hyperparameters of adaptive reweighting and regularization in an integrated manner based on the loss, and finally used the three hyperparameters to calculate the final ensemble meta-loss and update the base network, thereby improving its performance on biased distribution datasets. Experimental results demonstrate that EMN achieves higher accuracy on CIFAR and OCTMNIST datasets compared to other methods, and the effectiveness of diffe-rent strategies are demonstrated through policy correlation analysis. [ABSTRACT FROM AUTHOR]

Multi-focus image fusion(MFIF) is an image enhancement method that combines the focused regions from different source images to form a fully sharp image. Currently, in the context of MFIF methods, there are two main challenges. First, traditional methods such as spatial domain approaches produce fusion images with high objective scores, but they suffer from strong defocus spread effects(DSE) and artifacts at the fusion boundaries. Second, deep learning methods lack a dataset generated from plenoptic cameras and require extensive manual parameter tuning, resulting in time-consuming training processes. To address these challenges, this paper proposed an unsupervised multi-focus image fusion method based on target image prior information. Firstly, it utilized the internal prior information of the source image itself and the external prior information of the initial fusion image generated by a spatial domain method as inputs for the G-Net and F-Net, respectively, both the G-Net and F-Net were components of the UNet-based deep image prior(DIP) network. Then, it introduced a reference mask generated by a spatial domain method to assist G-Net network for generating a guiding decision map. Finally, it used the decision map and the initial fusion image to jointly optimize the F-Net, producing the final fusion image. It conducted validation experiments on the Lytro dataset with real reference images and the MFFW dataset with strong DSE exhibiting in the fusion boundaries, and employed five widely used objective metrics for performance evaluation. The experimental results demonstrate that the proposed method significantly reduces the number of optimization iterations, and outperforms eight state-of-the-art MFIF approaches in terms of the subjective and objective performance evaluation, and especially shows superior performance on the datasets with strong DSE exhibiting in the fusion boundaries. [ABSTRACT FROM AUTHOR]

The currently popular image inpainting methods based on deep neural network typically employ large receptive field feature extractors. However, when restoring local patterns and textures, they often generate artifacts or distorted textures, thus failing to recover the overall semantic and visual structure of the image. To address this issue, this paper proposed a novel image inpainting method, called ORFNet, which combined coarse and fine inpainting by employing an optimized receptive field strategy. Initially, it obtained a coarse inpainting result by using a generative adversarial network with a large receptive field. Subsequently, it used a model with a small receptive field to refine local texture details. Finally, it performed a global refinement inpainting by using an encoder-decoder network based on attention mechanisms. Validation on the CelebA, Paris StreetView, and Places2 datasets demonstrates that ORFNet outperforms existing representative inpainting methods. It leads to 1.98 dB increase in PSNR and 2.49% improvement in SSIM, along with average 2.4% reduction in LPIPS. Experimental results confirm the effectiveness of the proposed image inpainting method, showcasing superior performance across various receptive field settings and achieving more realistic and natural visual outcome. [ABSTRACT FROM AUTHOR]

In medical image segmentation tasks, the domain shift problem affects the performance of trained segmentation models in the unseen domain. Therefore, improving model generalization is crucial for the practical application of intelligent models for medical images. Representation learning is currently one of the dominant methods for solving domain generalization problems, mostly using image-level loss and consistency loss to supervise image generation. However, it is not sensitive enough to the deviation of small morphological features of medical images, which can lead to unclear edges of the generated images and affect the subsequent learning of the model. In order to improve the generalization of the model, this paper proposed a semi-supervised feature-level loss and learnable noise domain generalization(FLLN-DG) method for medical image segmentation. Firstly, the introduction of feature level loss improved the problem of unclear boundaries of the generated images. Secondly, the introduction of the learnable noise components further increased the data diversity and improved the model generalization. Compared with the baseline model, FLLN-DG improves the performance in the unseen domain by 2% to 4%, which demonstrates the effectiveness of to feature-level loss and to learnable noise components. FLLN-DG also has the best generalization and segmentation results compared to typical generalization models such as nnUNet, SDNet+AUG, LDDG, SAML and Meta. [ABSTRACT FROM AUTHOR]

In response to the issues of inadequate feature representation in existing face anti-spoofing algorithms and poor cross-dataset generalization performance, this paper proposed a face anti-spoofing method based on content-style enhancement and feature embedding optimization. Firstly, this method utilized a ResNet-18 encoder to extract generic features from multiple source domains, and then subjected to separation through two self-adaptive modules with different attention mechanisms, enhancing the representation of global content features and local style features. Secondly, based on the AdaIN algorithm, it organically fused content features with style features, further improving the feature representation, and the fused features were subsequently input to specific classifiers and domain discriminators for adversarial training. Finally, by employing average negative samples and semi-hard sample triplet mining to optimize feature embeddings, effectively striking a balance between intra-class cohesion and inter-class discrimination, better capturing the boundaries between genuine and spoofed categories. The proposed method was trained and tested on four benchmark datasets, suchas CASIA-FASD, REPLAY-ATTACK, MSU-MFSD, and OULU-NPU. It achieves accuracy of 6.33%, 12.05%, 8.38% and 10.59% respectively, which are superior to existing algorithms. This indicates that the proposed method can significantly improve the generalization performance of face live detection models in cross-dataset testing. [ABSTRACT FROM AUTHOR]

AT is a powerful means to defend against adversarial attacks. However, currently available methods often struggle to strike a balance between training efficiency and adversarial robustness. Some methods increase training efficiency but decrease adversarial robustness, while others do the opposite. To achieve the best trade-off, this paper proposed a joint adversa-rial training method based on a learnable attack step size(FGSM-LASS). This method included a prediction model and a target model. The prediction model predicted an attack step size for each example, which replaced the fixed-size attack step size using in the FGSM algorithm. Subsequently, the improved FGSM algorithm feeded both the target model parameters and original examples to generate adversarial examples. Finally, the prediction model and the target model perform joint adversarial training using these adversarial examples. Compared to the five most recent methods, FGSM-LASS was six times faster than LAS-AT, which was the best performing method in terms of robustness, with only 1% decrease in robustness. It was 3% more robust than ATAS, which was comparable in speed. Extensive experimental results fully demonstrate that FGSM-LASS outperforms current methods in the trade-off between training speed and adversarial robustness. [ABSTRACT FROM AUTHOR]

In order to address the problem of rapid increase of data throughput in fifth-generation wireless communication networks, mobile edge caching has become a useful solution. It can reduce the burden on the backhaul link and core network, cut down service latency by storing network content on edge devices. So far, most edge caching solutions have mainly focused on optimizing cooperative content caching, and ignored the efficiency of content transmission. This paper studied cooperative edge caching and wireless bandwidth allocation problems in ultra-dense networks, calculated the overall similarity between the base stations by using cosine similarity and Gaussian similarity, and grouped the small base stations according to total similarity in the network. The caching and radio bandwidth allocation problems were modeled as a long-term mixed-integer non-linear programming(LT-MINLP). Then, the cooperative edge caching and wireless bandwidth allocation problem were transformed into a constrained Markov decision process. Finally, it proposed cooperative edge caching and radio resource allocation scheme by using the DDPG model. And it proposed deep reinforcement learning based-edge content cooperative caching and bandwidth allocation algorithm CBDDPG. The proposed base station group strategy increased the file sharing opportunity among base stations, the cache scheme of the proposed CBDDPG algorithm used DDPG dual-network mechanism, which could better capture the regularity of user requests and optimize cache deployment. The proposed CBDDPG algorithm was compared to three baseline algorithms, sach as RBDDPG, LCCS and CB-TS in experiments. Experimental results show that the proposed strategy can effectively enhance the content cache hit ratio, reduce the delay of content delivery and improve the user experience. [ABSTRACT FROM AUTHOR]

Aiming at the problem of uneven inter-domain delay in slicing in multi-domain networks, this paper proposed an end-to-end dynamic cooperative slicing method based on inter-domain delay game. This method used game theory methods to allocate end-to-end delay constraints to different network domains, and deployed slices within the domain to obtain correspon-ding game benefits. It used the DDPG algorithm to continuously update the game strategy, and finally obtained the optimal delay allocation ratio and slice deployment solution. Experiments show that compared with the traditional static allocation algorithm, the proposed algorithm has obvious advantages. Compared with the empirically iterative DSDP method and the DQN-SNAF algorithm, the IDGA algorithm increased the slice deployment success rate by about 8% and 3% respectively under 100 slice requests. At the same time, the node resource utilization rate increased by about 5.75% and 1.96%. There are also significant advantages in reducing deployment costs. [ABSTRACT FROM AUTHOR]

In order to improve the fairness of downlink non-orthogonal multiple access systems, this paper proposed a dual layer squeezing iterative power allocation method based on fairness improvement. Firstly, it designed a user channel splitting process based on user channel gain difference to facilitate dynamic adjustment of the degree of power squeezing. Secondly, it carried out the inner power squeezing iteration from strong users to weak users, and set the squeezing factor and additional squeezing factor according to the channel splitting situation, so that the weak users at the channel splitting point had a stronger squeezing degree towards the strong users. Thirdly, the reallocation and resqueeze of the inner power through outer iteration could further improve the fairness of power distribution, while maintaining the total power unchanged. The simulation results show that compared with the fairness index without channel splitting, it can improve the fairness with channel splitting by 25%; Compared to the situation that the squeezing factor and additional squeezing factor are only adjusted without the number of inner and outer iterations adjustment, it improves the fairness index performance by 53.71%. Meanwhile, the fairness of proposed algorithm is better than the waterfilling algorithm and gained a good fairness effect. [ABSTRACT FROM AUTHOR]

The endoscopic defogging algorithm has a wide range of applications in the medical field, providing clear and real-time images for clinical doctors. Although significant progress has been made in defogging technology, the complexity of defogging algorithms is relatively high, making hardware implementation difficult in complex situations such as endoscopes. In order to achieve real-time defogging of endoscopes on hardware, this paper improved the dark channel prior algorithm to reduce hardware resource consumption and time complexity. The improved algorithm selected suitable atmospheric light intensity estimation values, transmittance compensation values for hardware, and used a pipeline structure to achieve the processing of foggy images. The hardware circuit of the algorithm was implemented using Xilinx's ZYNQ7020, with a real-time processing resolution of 640×480 video images, with a speed of up to 260 fps, consume only 1.28 K LUT, and consume 619 registers. The experimental results show that compared to traditional algorithms, the improved algorithm has the characteristics of fast processing speed, low power consumption, and strong portability, meeting the requirements of real-time video processing for endoscopes. [ABSTRACT FROM AUTHOR]

The task scheduling of e-government cloud center has always been a complex problem. Most existing task scheduling solutions rely on expert knowledge and are not versatile enough to deal with dynamic cloud environment, which often leads to low resource utilization and degradation of quality-of-service, resulting in longer makespan. To address this issue, this paper proposed a deep reinforcement learning(DRL) scheduling algorithm based on the actor-critic(A2C) mechanism. Firstly, the actor network parameterized the policy and chose scheduling actions based on the current system state, while the critic network assigned scores to the current system state. Then, it updated the actor policy network using gradient ascent, utilizing the scores from the critic network to determine the effectiveness of actions. Finally, it conducted simulation experiments using real data from production datacenters. The results show that this method can improve resource utilization in cloud datacenters and reduce the makespan in comparison to the classic policy gradient algorithm and five commonly used heuristic task scheduling methods. This evidence suggests that the proposed method is superiorly adapted for the dynamic e-government clouds. [ABSTRACT FROM AUTHOR]

Predictive business process monitoring(PBPM) represents a vital research field within BPM that aims to accurately predict future behavioral events. At present, deep learning methods are widely used in PBPM research. However, most of these methods consider only a single event-control flow perspective and do not fuse the attribute-data flow perspective for process prediction. To address this issue, this paper proposed a method called the fusion multi-perspective(FMP) framework based on a two-layer BERT neural network. Firstly, the first layer of BERT was used to learn attribute-data flow information. Subsequently, the second layer of BERT learnt event-behavior control flow information. Finally, the FMP framework combined data flow and control flow to achieve multi-perspective process prediction. Experimental results on real event logs demonstrate that, compared to other research methods, the FPM framework yields higher accuracy in predicting the next event activity. This validates that the FPM framework, which merges multi-perspective views of processes, enables a more comprehensive and in-depth analysis of complex process behaviors while enhancing predictive performance. [ABSTRACT FROM AUTHOR]

Focused on the problem of current software remaining life prediction methods ignore the life information contained among multiple performance indicators, this paper proposed a remaining life prediction method of Web software system based on the Transformer model with multiple performance indicators(TransMP). Firstly, it constructed an accelerated aging experimental platform for a memory fault type Web system, and created a dataset containing performance indicators such as memory usage, response time, and throughput. Secondly, considering the differences in aging characteristic information contained in different performance indicators, it constructed the TransMP model consisting of multiple encoders-decoders. The performance indicator data was separately input into the memory indicator encoder, response time encoder, and throughput encoder to extract aging characteristic information, then it introduced a feature fusion layer for information fusion. Finally, it input the fused information into the decoder composed of a mask attention-multi-head attention structure to predict the remaining life when the system reached an aging threshold. The experimental results indicate that the remaining life prediction method of the Web system, compared to the SALSTM method, reduces the root mean square error by 12.0%, 17.3% and 13.2%, and decreases the mean absolute error by 13.3%, 21.0% and 10.4%, demonstrating the effectiveness of this method. [ABSTRACT FROM AUTHOR]

TTA aims to make the trained neural network model adapt to the test data distribution at test time using unlabeled test data. Existing TTA methods mainly consider adaptation in a single or multiple static environments. However, in nonstationary environments, the test data distribution changes continuously over time, which leads to the instability of previous TTA methods. Therefore, this paper proposed a test-time adaptation algorithm(RRSTA) based on robust and reliable symmetric cross entropy. Firstly, in order to improve the robustness to noise distribution changes and alleviate catastrophic forgetting, it proposed a symmetric cross entropy based on the mean teacher model, which encouraged the model to predict correctly and punished the wrong prediction. Secondly, in order to improve the robustness to different noise samples, it proposed a dual-stream perturbation technique, which guided the weak-to-strong perturbation view of the student model through the strong view of the teacher model. Finally, it proposed a reliable entropy minimization strategy to prevent the drastic change of parameters and stabilize adaptation. Extensive experiments and ablation studies on CIFAR10C and CIFAR100C confirm the effectiveness of the proposed method. Compared with the unadapted models, the error rate is significantly reduced by 26.13% and 14.69%, and it is significantly better than the second-best method. [ABSTRACT FROM AUTHOR]

Addressing the issues of dynamic complicated constraints and the lack of visualization in automatic guided vehicle(AGV) scheduling at automated container terminals, this paper utilized the AnyLogic platform to construct a computational experiment model of an automated container terminal within the conceptual framework of computational logistics. It proposed two aspects of improvement for the AGV scheduling strategies within the model. Firstly, it employed the ILOG CPLEX to the AnyLogic system to solve dynamic complicated combinatorial optimization problem. Secondly, it devised five problem-oriented customized AGV scheduling strategies based on computational logistics. Due to the inhe-rent limitations of the CPLEX solver, the model and experimental data primarily focused on the import containers of the automated container terminal. The computational experiments demonstrate that the customized sche-duling strategies effectively complement the limitations of the CPLEX and the AnyLogic. In comparison to the original system policies, the customized strategies can better realize the production scheduling of AGVs for specific loading and unloading needs, while meeting efficient scheduling requirements. [ABSTRACT FROM AUTHOR]

The performance of most existing evolutionary algorithms tends to decrease as the dimension of the decision variables increases for solving large-scale optimization problems. The Pareto solution set of a multi-objective optimization is a low dimensional manifold in the decision space, whose dimension is much smaller than that of the decision variables space. Accordingly, this paper proposed a multi-objective evolutionary algorithm based on dimensionality reduction of decision variables to solve large-scale sparse multi-objective optimization problems. It preserved local neighborhood information in the original decision variables space by using locality preserving projections, and designed an archive set to train the non-dominated solutions as much as possible to raise the accuracy of projection. The proposed algorithm was compared with four popular evolutionary multi-objective optimization algorithms on a series of test problems and practical application problems. Experimental results show that the proposed algorithm is effective in solving sparse multi-objective problems. Therefore, the reduction of independent variables can reduce the difficulty of solving the problem, improve the search efficiency of the algorithm, and have significant advantages in solving large-scale sparse multi-objective problems. [ABSTRACT FROM AUTHOR]

The researches on driver pathway identification in cancer rely on traditional biological experiments, which have the drawbacks of being time-consuming, labor-intensive and costly. This paper proposed a novel binary cancer driver pathway identification method called PEA-BLMWS(parental evolutionary algorithm-binary linear maximum weight sub-matrix). Firstly, it utilized the existing gene expression data to uncovered potential gene mutation data by comparing the differences in expression levels between normal and mutated genes. Secondly, it incorporated protein-protein interaction network data to construct an improved binary linear maximum weight sub-matrix model. Finally, it proposed a parental evolutionary algorithm to solve this matrix model. Experimental results on the GBM(glioblastoma) and OVCA(ovarian cancer) datasets show that compared to other advanced identification methods such as Dendrix, CCA-NMWS and CGP-NCM, the gene set identified by PEA-BLMWS has more genes enriched in known signaling pathways, and genes not enriched in signaling pathways are also closely related to the occurrence of cancer. Therefore, this identification method can serve as an effective tool for driving pathway identification. [ABSTRACT FROM AUTHOR]

In order to enhance the accuracy of vehicle positioning in the Internet of Vehicles, this paper proposed a collaborative positioning method for the Internet of Vehicles based on fusing vehicle-mounted radar ranging information with GNSS information. This method established a mathematical model using the maximum likelihood estimation strategy, which essentially solved a nonlinear optimization problem. It simplified the problem as a quadratic programming problem with multiple quadratic constraints, presented a semi-definite relaxation method to efficiently approximate the original problem, and used the eigenva-lue decomposition method to further improve the approximate solution. Simulation results demonstrate that the accuracy of the collaborative positioning achieved by this information fusion method significantly improves compared to the linearized weighted least squares method. It can also achieve the positioning accuracy of BP(back propagation) neural network localization method based on a large dataset, without the need for pre-training the model, enabling high-precision real-time positioning. [ABSTRACT FROM AUTHOR]