Abstract:Aiming at the difficulty in manually designing portrait paper-cuts, this study employs the generative adversarial network (GAN) to generate high-quality portrait paper-cuts for the first time. Based on the artistic characteristics of portrait paper-cuts, an improved network based on CycleGAN is proposed. 1) The CBAM attention module is introduced into the CycleGAN generator to enhance the feature extraction of the network. 2) The local discriminator for key facial regions such as nose, eyes, and lips is introduced to improve the generation effect of the above areas in generated portrait paper-cuts. 3) A new loss function is designed based on image edge information and SSIM, which will be adopted to replace the original forward cycle-consistency loss of CycleGAN and eliminate the shadows in the portrait paper-cuts. Compared with other automatic generation methods of portrait paper-cuts, the proposed method can quickly generate paper-cuts featuring high similarity to the original human face, continuous and smooth lines, and aesthetic beauty. Additionally, this study also puts forward a post-processing method of portrait paper-cut connectivity to make the obtained results more consistent with the overall connectivity of traditional Chinese paper-cuts.

Abstract:In the field of image compression perceptual reconstruction, high-quality images are reconstructed with high similarity to the original image, and details are retained to eliminate artifacts through effective image prior information reconstruction. Thus, aiming at the K-space data with insufficient sampling, based on the classic CNN algorithm CBDNet algorithm, this study adopts the method to combine the advantages of fusing deep learning prior information and traditional image restoration. Meanwhile, a hybrid reconstruction algorithm based on prior denoising of deep neural network and compressed sensing algorithm of BM3D block is studied. The algorithm employs an interactive method to train a multi-scale residual network to suppress noise levels and combines deep learning with the multi-scale matching of traditional blocks to extract image feature data at different scales through optimal selection, thus suppressing artifacts and quickly reconstructing high-quality MRI. The experimental results show that deep learning combined with BM3D can reduce artifacts and retain details in MR image reconstruction, enhancing the reconstruction effect. Additionally, the computational complexity of the algorithm is not much more than that of the single algorithm by the GPU accelerated operation. It can be seen that the hybrid MRI based on convolution blind denoising has a better effect.

Abstract:Helpfulness prediction task of online reviews is significant in the contemporary e-commerce environment. It aims to evaluate the helpfulness of online reviews and then highlight the reviews more helpful to future consumers, thereby improving the consumers’ efficiency in obtaining information. This study concentrates on the new multidimensional scoring system emerging on various online platforms in recent years, and tries to study the influence of aspect ratings given by users in the system on the helpfulness of online reviews. To accomplish the helpfulness prediction task, it puts forward a multi-level neural network model HORA that considers all three components of review texts, overall ratings, and aspect ratings, as well as their interconnections. The experimental results on two real-world datasets show that HORA outperforms the present baseline models in terms of MAE and RMSE and exhibits good robustness. This indicates the significance of aspect ratings for the helpfulness awareness of users’ online reviews.

Abstract:The research on natural language to SQL (NL2SQL) has high application value. With the maturity of deep learning technology, increasingly more researchers have begun to apply deep learning technology to NL2SQL tasks. This study reviews the research status of NL2SQL in English and Chinese fields and summarizes the datasets and models published by year. Additionally, it compares the characteristics of the four major Chinese NL2SQL datasets and expounds on the basic framework of NL2SQL tasks based on deep learning and typical models for simple single-table problems and complex cross-table problems in Chinese NL2SQL fields. Finally, the commonly adopted model evaluation methods are introduced, and future research directions are put forward.

Abstract:To solve the low learning efficiency and slow convergence due to the complex relationship among intelligent agents in multi-agent reinforcement learning, this study proposes a two-level attention mechanism based on MADDPG-Attention. The mechanism adds soft and hard attention mechanisms to the Critic network of the MADDPG algorithm and learns the learnable experience among intelligent agents through the attention mechanism to improve the mutual learning efficiency of the agents. Since the single-level soft attention mechanism assigns learning weights to completely irrelevant intelligent agents, hard attention is employed to determine the necessity of learning between two intelligent agents, and the agents with irrelevant information are cut. Then soft attention is adopted to determine the importance of learning between two intelligent agents, and the learning weights are assigned according to the importance distribution to learn from the agents with available experience. Meanwhile, tests on a collaborative navigation environment with multi-agent particles show that the MADDPG-Attention algorithm has a clearer understanding of complex relationships and achieves a success rate of more than 90% in all three environments, which improves the learning efficiency and accelerates the convergence rate.

Abstract:At present, breast cancer, with the highest annual incidence, has replaced lung cancer, and the target detection technology based on deep learning can automatically detect lesions on non-invasive imaging such as mammography X-ray, breast ultrasound, and breast magnetic resonance imaging (MRI), and it has become the preferred way for adjuvant diagnosis of breast cancer. You only look once (YOLO) series algorithms are object detection algorithms based on deep learning, and classical YOLO algorithms have certain advantages in speed and accuracy and are widely used in computer vision fields. The latest YOLO algorithm is the state of the art (SOTA) model in the field of computer vision, and how to use YOLO series algorithms to improve the speed and accuracy of breast cancer detection has become one of the focus of researchers. On this basis, this study introduces the principle of the classical YOLO series algorithms, sorts out the application status of the classical YOLO series algorithms in breast cancer image detection, summarizes the existing problems, and looks forward to the further application of the YOLO series algorithms in breast cancer detection.

Abstract:In traditional control systems, people rely on employing devices such as handles and joysticks to achieve human-machine interaction with external devices, which is a challenge for patients with movement disorders. Meanwhile, brain-computer interface (BCI) technology can convert EEG into control commands for external devices through the brain loop, allowing these patients to directly control external devices by their brain’s “consciousness”. This study proposes an autonomous driving system of intelligent car based on multimodal BCI to integrate the subjects’ EEG, electro-oculography, and gyroscope signals to control the car. EEG is used for controlling the car speed, electrooculography for controlling the start and stop of the car, and gyroscope signals for controlling the car steering. Additionally, computer vision technology is combined to add autonomous driving function for the intelligent car, making control more intelligent. The experiments show that the average accuracy rate of ten subjects utilizing the system to control the car is 92.47%, with an average response time of 1.55 s and an average information transmission rate of 55.94 bit/min, which indicates the effectiveness and efficiency of the control system. Meanwhile, multiple comparative experiments for verification are set up to verify the car’s autonomous driving function. The experimental results show that compared with manual driving, although the autonomous driving system has disadvantages in controlling the car speed, it has better performance advantages in accuracy and stability. This proves that this system can provide better control experience for the disabled, and has broad application prospects in brain control and autonomous driving.

Abstract:To solve the scarcity of text classification algorithms in finance and the inability of existing algorithms to adequately extract word-to-word relations, long-distance dependency, and deep feature information in texts, this study proposes a text depth relationship extraction algorithm based on improved convolutional self-attention model. The algorithm introduces self-attention in a modified deep pyramidal convolutional neural network (DPCNN) and builds a text classification model jointly with bi-directional gated neural network (BiGRU) module to solve the problem of extracting long-distance dependency feature information and word-to-word relationship feature information for long texts in finance. Then the joint extraction function of deep feature information and contextual semantic information in texts is realized. Experiments on THUCNews short text and long text datasets show that the proposed method has significant improvement in evaluation indexes compared with BERT and other methods. The comparison experiments on the dataset of homemade financial long texts show that the accuracy and F1 value of the algorithm model are higher compared with other models. A series of experiments demonstrate that the algorithmic model can perform the classification task against financial long texts more accurately.

Abstract:Predicting the trend of inlet valve temperature changes provides significant references for the operating status of valve cooling systems. Since the traditional methods have problems such as a large time span of data collection and sensor deviation, this study proposes a Robust-InTemp prediction model for inlet valve temperature based on adversarial perturbation and local information enhancement. Specifically, Robust-InTemp enhances the model’s generalization ability and noise resistance robustness by adding rule-based Gaussian noise to the original data and employing projected gradient descent (PGD) for adversarial training. Meanwhile, relative positional encoding, one-dimensional convolution, and gated linear units (GLUs) are introduced to enhance the model’s ability to learn local features, thus improving prediction accuracy. Experimental results show that compared to various benchmark models, Robust-InTemp has clear advantages in predictive performance and anti-interference ability. Additionally, further ablation experiments validate the effectiveness of each component in the model.

Abstract:In multi-user and multi-task scenarios, using traditional decision algorithms to make computation offloading decisions for upcoming tasks in a short period can no longer meet users’ requirements for decision-making efficiency and resource utilization. Therefore, some studies have proposed deep reinforcement learning algorithms for offloading decisions to cater to various scenarios. However, most of these algorithms only consider the offloading first strategy, which leaves user equipment (UE) idle. This study improves the resource utilization of mobile edge computing (MEC) servers and UE and reduces the error rate of computation offloading. It proposes a decision offloading model combining local first and improved twin delayed deep deterministic policy gradient (TD3) algorithm and designs a simulation experiment. The experimental results show that the model can indeed improve the resource utilization of MEC servers and UE and reduce the error rate.

Abstract:Label noise can greatly reduce the performance of deep network models. To address this problem, this study proposes a contrastive learning-based label noisy image classification method. The method includes an adaptive threshold, contrastive learning module, and class prototype-based label denoising module. Firstly, the robust features of the image are extracted by maximizing the similarity between two augmented views of the same image using contrastive learning. Then, a novel adaptive threshold filtering training sample is used to dynamically adjust the threshold based on the learning status of each class during model training. Finally, a class prototype-based label denoising module is introduced to update pseudo-labels by calculating the similarity between sample feature vectors and prototype vectors, thus avoiding the influence of label noise. Comparative experiments are conducted on the publicly available datasets CIFAR-10 and CIFAR-100 and the real dataset ANIMAL10. The experimental results show that under the condition of artificially synthesized noise, the proposed method outperforms conventional methods. By updating pseudo-labels based on the similarity between the robust feature vector of the image and various prototype vectors, the negative impact of noisy labels is reduced, and the anti-noise ability of the model is improved to certain extent, verifying the effectiveness of the proposed model.

Abstract:Regarding the challenge of handling nested medical entities in Chinese electronic medical records, this study proposes a knowledge-enhanced named entity recognition model for Chinese electronic medical records called ERBEGP based on the RoBERTa-wwm-ext-large pre-trained model. The comprehensive word masking strategy employed by the RoBERTa-wwm-ext-large model can obtain semantic representations at the word level, which is more suitable for Chinese texts. First, the model learns a significant number of medical entity nouns by integrating knowledge graphs, further improving entity recognition accuracy in electronic medical records. Then, the contextual semantic information within the records can be better captured through BiLSTM encoding of the input sequence of medical records. Finally, the efficient GlobalPointer (EGP) model is adopted to simultaneously consider the features of both the head and tail of entities to predict nested entities, addressing the challenge of handling nested entities in named entity recognition tasks of Chinese electronic medical records. The effectiveness of the ERBEGP model is demonstrated by yielding better recognition results on the four datasets within CBLUE.

Abstract:Many skin cancer diseases have obvious early symptoms. Currently, the diagnosis of skin cancer mainly relies on medical workers with professional knowledge, bringing the problems such as long time consumption and low reusability. In response to these problems, a lightweight skin disease recognition model based on improved MobileNetV3-Small is proposed in this study. Firstly, a CaCo attention module based on coordinate attention (CA) mechanism is proposed, Secondly, for the uneven distribution of the samples of skin-cancer datasets, a combination of multiple loss functions is proposed to enhance the learning ability of the model for cases with few samples. The improved MobileNetV3-CaCo model has an accuracy, balance accuracy, and model parameter quantity of 93.39%, 86.35%, and 2.29M, respectively, thus ideal recognition results are achieved.

Abstract:To solve the poor visual experience caused by the uncoupled ship-wave motion in existing navigation simulators, this study develops a set of visual simulation systems for real-time interaction between ship and wave motion. Firstly, the wave motion scene is built by wave spectrum modeling technology. Then, based on building the ship force model, the ship’s response to wave force is realized by collision detection between the ship and the water surface to calculate the real-time response attitude of the ship to wave motion. Meanwhile, the wave formula is adopted to calculate the water wave generated by the collision between the ship and the water body and its diffusion and enhance the realism of the simulation system. Compared with traditional navigation simulators, the navigation simulator coupled with ship-wave motion can provide more realistic visual and sports experience in high level sea conditions. The results indicate strong visual realism and sound real-time interaction between ship and wave, with a sound simulation effect on the navigation in bad navigation conditions.

Abstract:To improve the seal detection efficiency of threaded oil casing gas, this study proposes an automatic classification network, NAFENet, for threaded torque curves based on global attention feature fusion. Specifically, NAFENet extends the convolutional structure of EfficientNet-B0 to 11 layers to obtain EfficientNet-B11 and enhance the model expressiveness. Meanwhile, the modules based on non-local global attention and attentional feature fusion (AFF) are built in each MBConv convolutional layer to help the model acquire more global information in the curve images and improve the feature extraction ability. The experimental results show that compared with EfficientNet-B0, the parameter number of NAFENet is slightly increased with improved curve identification accuracy, and the model accuracy reaches 92.87% on the homemade UBT_Curve dataset.

Abstract:Graph partitioning algorithm is part of distributed graph computing system. It divides a graph into several subgraphs to run in the distributed system and assigns the vertex and edge data and computing tasks on the subgraphs to each partition. Heterogeneous graph is a kind of graph widely existing in the real world. It is a graph with multiple vertex types or edge types. During the calculation of heterogeneous graphs, the existing graph partition algorithms do not consider the following problems. In graph calculation, different vertex and edge types may carry different data amounts. Meanwhile, different vertex and edge types may adopt different processing algorithms, with various calculation time. Aiming at the shortcomings of the existing graph partitioning methods, this study proposes an online graph partitioning algorithm for heterogeneous graphs, OGP-HG algorithm. Additionally, the existing GraphX graph computing engine is improved to implement the proposed algorithm in the improved graph computing engine. The proposed OGP-HG algorithm calculates the load balance score of vertices divided into different partitions and the data balance score of edges divided into different partitions, thus obtaining the division results of balancing the load and memory occupation of heterogeneous graphs. Experiments show that compared with traditional graph partitioning algorithms, this algorithm improves the computing efficiency of heterogeneous graphs by 1.05–1.4 times.

Abstract:To address the low feasibility of human pose estimation algorithms and low accuracy of jump rope counting based on pose estimation, this study proposes a jump rope counting algorithm based on a lightweight human pose estimation network. The algorithm first inputs a jump rope video, then extracts keyframe images by inter-frame difference method, and feeds them into the human pose estimation network for key joint point detection. To improve the detection accuracy of the lightweight network, the study builds an optimized LitePose detection model, which employs adaptive perception decoding to optimize the decoding part in the model and reduce quantization errors. Furthermore, a Kalman filter is adopted to smooth and denoise the coordinate data, reducing coordinate jitter errors. Finally, jump rope counting is determined based on the changes in key-point coordinates. Experimental results demonstrate that, in the same image resolution and environmental conditions, the proposed algorithm employing the optimized LitePose-S network model does not increase the parameter size and computational complexity of the model but improves network detection accuracy by 0.7% compared with other comparison networks. Meanwhile, the average error rate of this algorithm in jump rope counting can reach a minimum of 1.00%. The algorithm effectively determines the takeoff and landing of the human body by the results of human pose estimation and yields counting results.

Abstract:To solve the slow convergence, poor stability, and proneness to fall into local extremes of traditional path planning algorithms, this study proposes a vehicle path planning method based on a gradient statistical mutation quantum genetic algorithm. Firstly, based on the dynamic adjustment of the rotation angle step by the chromosome fitness value, the idea of gradient descent is introduced to improve the adjustment strategy of the quantum rotation gate. According to the statistical characteristics of chromosome variation trend, a mutation operator based on gradient statistics is designed to realize mutation operation, and an adaptive mutation strategy based on Qubit probability density is put forward. Then the vehicle path planning model is built with the shortest path as the index. Finally, the effectiveness of the improved algorithm in vehicle path planning is verified by simulation experiments. Compared with other optimization algorithms, the proposed algorithm has a shorter path and better search stability to avoid the algorithm from falling into the local optimum.

Abstract:In gear train design, traditional algorithms exhibit drawbacks such as computational complexity and low accuracy. The seagull optimization algorithm (SOA) benefits from its simple algorithmic principle, strong universality, and few parameters, and is now commonly used in engineering design problems. However, the standard SOA is prone to problems such as low optimization accuracy and slow search speed. This study proposes a hybrid strategy improved seagull optimization algorithm (WLSOA). Firstly, it utilizes a nonlinear descent strategy to enhance the exploration and development capabilities of the SOA and improve optimization accuracy. Secondly, the adaptive weight balancing of global and local search capabilities and the addition of Levy flight steps to perturb the current optimal solution are introduced to improve the ability of the algorithm to jump out of the local optimal value. The performance of WLSOA is then explored through simulation experiments on 9 classic test functions, using WLSOA, golden sine algorithm, whale optimization algorithm, particle swarm optimization algorithm, traditional seagull optimization algorithm, and the newly proposed improved seagull optimization algorithm. The results show that WLSOA has higher optimization accuracy and faster convergence speed than the other six algorithms. Finally, in gear train design, a comparison with 13 other common swarm intelligence algorithms reveals that WLSOA has a better solving ability than other algorithms.

Abstract:With the development of intelligent transportation, a large amount of vehicle trajectory data is collected and stored. However, the trajectory data always has anomalous trajectory point data, seriously affecting the accuracy and effectiveness of subsequent trajectory data analysis. This study finds a class of implicit positional anomaly trajectory data that is difficult to be detected by traditional detection methods based on movement feature thresholds but plays a vital role in trajectory data analysis. To this end, this study proposes a method to detect the implicit anomalous trajectory data based on floating grid and clustering method. The parallelization method of data is realized by taking the trajectory data of some cabs in Xi’an as an example. The experimental results show that the data recall and accuracy of the proposed method to detect the hidden location anomaly could reach 0.90, and the F1-score is in the range of 0.88–0.91. The detection of such implicit anomalous trajectory data is beneficial to subsequent analysis and application of spatio-temporal trajectory data.

Abstract:AIS data refers to the vessel’s motion trajectory information obtained through the AIS system. Mining AIS data can provide insights into the vessel’s motion patterns, navigation routes, docking locations, etc. However, outliers generated during the AIS data collection can have a negative effect on clustering and other tasks. Therefore, outlier detection on AIS data before mining is necessary. However, when there are a large number of outliers in AIS trajectory data, a significant decrease occurs in the accuracy of most outlier detection algorithms. To address this issue, this study proposes a trajectory outlier detection based on center shift (CSOD). The CSOD algorithm encourages data points to move towards the center of their K-nearest neighbor (KNN) set, making each data point closer to typical data and effectively eliminating the influence of outliers on clustering. To validate the effectiveness of the proposed algorithm, the study conducts comparative experiments between the CSOD algorithm and several classical outlier detection algorithms using the AIS fishing vessel trajectory dataset in the Zhejiang sea area. The experimental results demonstrate that the CSOD algorithm outperforms the other algorithms in terms of overall performance.

Abstract:The core of penetration testing is to discover penetration paths, but not all penetration paths can be successful. Therefore, the optimal penetration path needs to be chosen based on the current system environment. In this context, firstly, this study models the environment as a Markov decision process (MDP) graph based on the attack graph and uses a value iteration algorithm to find the optimal penetration path. Secondly, a new replanning algorithm is proposed to deal with the failure of penetration actions in the MDP graph and find the optimal penetration path again. Finally, in view of the existence of multiple attack targets in the penetration testing process, this study proposes a multi-objective global optimal penetration path algorithm for MDP graphs. Experimentally, the proposed algorithm shows higher efficiency and stability in replanning tasks and is effective in multi-objective tasks, which can prevent unnecessary penetration actions from being executed.

Abstract:To address the problem that the solution accuracy of the sparrow search algorithm (SSA) depends on the population at the better location and is easily trapped in the local optimum, this study proposes an improved sparrow search algorithm (ISSA). The algorithm firstly proposes a normal shift strategy to shift the population with the center of gravity as the guide to achieve the decay of the normal distribution of the moving energy and effectively improve the exploration ability of the population for local search. Secondly, it introduces a dynamic sinusoidal perturbation strategy to achieve the two-way demands of the discoverer for the early search step and the late fast convergence through the scaling factor. Then, a backward learning mechanism is added for the poorly positioned early warners in the sparrow population to generate the backward solution of the perturbation with their current position, which is helpful to expand the search step and enable the algorithm to jump out of the local optimum. Finally, six test functions are randomly selected and compared with other similar algorithms, and the experimental results verify the effectiveness of the ISSA algorithm.

Abstract:In recent years, remote sensing images have been widely employed in a series of work such as environmental monitoring. However, the images observed by satellite sensors often have low resolution, which is difficult to meet in-depth research needs. Super resolution (SR) aims to improve image resolution and provides finer spatial details, perfectly compensating for the weaknesses of satellite imagery. Therefore, a back-projection attention network (BPAN) is proposed for SR reconstruction of remote sensing images. The BPAN is composed of the back-projection network and the initial residual attention block. In the back projection network, the iterative error feedback mechanism is adopted to calculate the upper and lower projection errors to guide image reconstruction. In the initial residual attention block, the initial module is introduced to integrate local multilevel features to provide more information for reconstructing detailed textures to focus on the importance of the module to learn different spatial regions adaptively and promote high-frequency information recovery. To evaluate the effectiveness of this method, this study conducts a large number of experiments on AID datasets. The results show that the proposed network model improves the reconstruction performance of traditional deep networks and has significant improvements in visual effects and objective indicators.

Abstract:The quality of graph partitioning greatly affects the communication overhead and load balance among computers, which is crucial for the performance of large-scale parallel graph computation. However, as the scale of graph data continues to increase, the execution time and memory overhead of graph partitioning algorithms have become inevitable. Therefore, it is necessary to study how to optimize the execution efficiency of graph partitioning algorithms. This study proposes a heuristic graph partitioning method based on weighted graph generation by breadth-first traversal, which introduces only a small amount of preprocessing time overhead while achieving lower communication overhead and better load balance. Experimental results show that our partitioning method reduces replication factors, lowers communication overhead, and only introduces a small amount of time overhead.

Abstract:The distribution of remaining oil forms is of great significance for the deep development of oil fields. This study proposes a form classification method of remaining oil based on deep learning to address the problems of scarce remaining oil data and the limited ability of traditional morphological parameter classification. In the data preprocessing stage, the method uses the multi-class data generation characteristics of the generative adversarial network (ACGAN) to enhance the data of the remaining oil image. It employs the VGG19 model as the backbone network to extract deep features that cannot be described by traditional morphological parameters and introduces the SENet attention mechanism to improve the model’s feature expression ability, making the final classification results more accurate. To verify the effectiveness, the proposed method is compared with traditional classification methods based on morphological parameters and other deep learning models, and it is evaluated through subjective visual and objective indicators. The results showed that the proposed method provides a more accurate classification.

Abstract:Interactive image segmentation is an important tool for pixel-level annotation and image editing. Most existing methods adopt two-stage prediction: first predicting a rough result, and then refining the previously predicted results in the second stage to obtain more accurate predictions. To ensure the viability of the network model under limited hardware resources, the same network is shared across the two stages. To better propagate labeled information to unlabeled areas, a similarity constraint propagation module is designed. Meanwhile, a simple prototype extraction module is used during training to make forward click vectors highly cohesive, accelerate network convergence, and remove them during inference. At the inference stage, the implementation of intention perception modules to capture details further improves prediction performance. Numerous experiments show that the method is most comparable to the most advanced methods on all popular benchmark tests, demonstrating its effectiveness.

Abstract:Aiming at the small target scale and low detection accuracy in traffic signal detection, this study proposes a traffic signal detection algorithm based on improved YOLOv5s. Firstly, a feature pyramid module RSN-BiFPN is constructed to fully integrate traffic signal features of different scales to reduce target missed detection and false detection. Secondly, a new feature fusion layer and prediction head are introduced to improve the perception performance of the network for small objects and enhance detection accuracy. Finally, the EIoU function is adopted to optimize the loss and accelerate network convergence. Experiments conducted on the public dataset S2TLD show that compared with the basic network, the precision rate of the proposed method is increased by 4.1% at 96.1%, the recall rate is 95.9% with an increase of 3%, and the average precision is increased by 1.9%, reaching 96.5%. Meanwhile, the improved algorithm achieves a faster detection speed of 22.7 frames per second. The proposed method can realize rapid and accurate detection of traffic lights and can be widely employed in the research on analyzing traffic lights.

Abstract:Neural machine translation technology can translate the semantic information of multiple languages automatically. Therefore, it has been applied to binary code similarity detection of cross-instruction set architecture successfully. When the sequences of assembly instructions are treated as sequences of textual tokens, the order of instructions is important. When binary basic block-level similarity detection is performed, the neural networks model instruction positions with position embeddings, but it failed to reflect the ordered relationships (e.g., adjacency or precedence) between instructions. To address this problem, this study uses a continuous function of instruction positions to model the global absolute positions and ordered relationships of assembly instructions, achieving the generalization of word order embeddings. Firstly, the source instruction set architecture (ISA) encoder is constructed by Transformer. Secondly, the target ISA encoder is trained by triplet loss, and the source ISA encoder is fine-tuned. Finally, the Euclidean distances between embedding vectors are mapped to [0,1], which are used as the similarity metrics between basic blocks. The experimental results on the public dataset MISA show that the evaluation metric P@1 of this study is 69.5%, which is 4.6% higher than the baseline method MIRROR.

Abstract:Automatic program repair techniques can realize automatic repair of software defects and employ test suites to evaluate repair patches. However, because of inadequate test suites, the patches passing the test suites may not repair the defects correctly, or even introduce new defects with ripple effects, which results in a large number of overfitting patches generated by automatic program repair. To this end, an overfitting patch identification method based on data flow analysis is proposed. This method firstly decomposes the patch modifications to the program into operations on variables, then adopts data flow analysis to identify the patch influence domain, and selects targeted coverage criteria to identify target coverage elements according to the domain. Finally, test paths are selected and test cases are generated to fully test the repair program to avoid the impact of repairing side effects. This study conducts evaluations on two datasets, and the experimental results show that the overfitting patch identification method based on data flow analysis can improve the correctness of automatic program repair.

Abstract:Many studies apply Transformer to time series prediction tasks. However, compared with other time series, motion trajectory data has kinematic uncertainty without obvious periodicity. To reduce noise interference and enhance trend modeling, this study proposes a target trajectory prediction method based on time-frequency domain information fusion and multi-scale adversarial training based on Transformer architecture. The wavelet decomposition is embedded into the network model to realize the adaptive filtering in the time-frequency domain, and then time-domain attention is integrated to encode the long-term trend characteristics of the observed trajectory more effectively. Meanwhile, the study designs a full convolution discriminator to further improve the prediction accuracy by learning multi-scale short-term micro motion representation of the sequence through adversarial training. A trajectory prediction dataset DT including 2D ship trajectory and 3D aircraft trajectory is established as a benchmark, and comparative experiments with Transformer, LogTrans, Informer, and other models are conducted. Experiment results show that the proposed method is superior to other models in the tasks of medium and long-term trajectory prediction.

Abstract:Prediction based on historical data has become essential in many fields, such as environmental management and urban transportation. Prediction accuracy plays a key role in practical production, scheduling, and other tasks. However, due to natural or human factors, some data exhibits high volatility and uncertainty, unable to fully achieve the potential of prediction models. Taking the sediment concentration prediction during the non-ice period as a case study, this study explores optimization methods for predicting high-volatility data. The results show that the feature selection optimization based on the Shapley additive explanations (SHAP), the data smoothing, and early-stage clustering can reduce prediction error of high-volatility data. The mean absolute error (MAE) decreases from 1.502 in the initial model to 0.194, and data smoothing shows the most significant optimization effect with a reduction of 76.51% in MAE. However, the increasing smoothing order results in poorer prediction results, which is because the subsequent rising exponentiation order correspondingly leads to an exponential increase in error. Additionally, employing clustering results as feature inputs can “guide” the parameter learning of multi-layer perceptron.

Abstract:In the booming autonomous driving technology, the results of pedestrian trajectory prediction often affect autonomous driving safety. Pedestrian trajectory prediction technology currently faces the problem of interaction with others when applied to practical scenarios, requiring consideration of social interaction and logical consistency during predicting trajectories. Therefore, this study proposes a pedestrian trajectory prediction method based on spatio-temporal graphs. This method employs graph attention networks to model pedestrian interactions in the scenarios and adopts a method of automatically generating positive and negative samples to reduce the collision rate of the output trajectory through contrastive learning, thus improving the safety and logical consistency of the output trajectory. Model training and testing are conducted on ETH and UCY datasets, and the results show that the proposed method reduces the collision rate and has better prediction accuracy than mainstream algorithms.