With the popularization of university information system applications and the increase in their usage frequency, teachers and students have higher requirements for data consistency, accuracy, timeliness, and completeness. The original data synchronization scheme using extensible markup language (XML) for data synchronization has the disadvantages of low synchronization efficiency and difficulty of expansion. The open-source tool, DataX, can complete data synchronization between various heterogeneous databases without damaging the source database. This study used DataX to improve the original data synchronization scheme and proposed different data synchronization schemes for various business requirements and application scenarios in the foundation of university postgraduate information system construction. At the same time, in view of the shortcomings of DataX in which only one read can do one write during the start-up and execution, the method where one read can do multiple writes was designed. The comparison experiment shows that the optimized scheme can improve data synchronization efficiency, has better scalability, and can meet the data synchronization requirements of universities.
Review-based recommendations are mainly based on the exploitation of textual information that reflects the characteristics of items and user preferences. However, most existing approaches overlook the influence of information from hidden strangers on the selection of reviews for the target user. However, information from strangers can more accurately measure the relative feelings of the user and provide a complement to the target user’s expression, leading to more refined user modeling. Recently, several studies have attempted to incorporate similar information from strangers but ignore the use of information regarding other strangers. In this study, we proposed a stranger collaborative review-based recommendation model to make effective use of information from strangers by improving accurate modeling and enriching user modeling. Specifically, for capturing potential user preferences elaborately, we first designed a collaborative stranger attention module considering the textual similarities and preference interactions between the target user and the hidden strangers implied by the reviews. We then developed a collaborative gating module to dynamically integrate information from strangers at the preference level based on the characteristics of the target user-item pair, effectively filtering preferences of strangers and enriching target user modeling. Finally, we applied a latent factor model to accomplish the recommendation task. Experimental results have demonstrated the superiority of our model compared to state-of-the-art methods on real-world datasets from various sources.
This study explores the multimodal understanding and reasoning for one-stage visual grounding. Existing one-stage methods extract visual feature maps and textual features separately, and then, multimodal reasoning is performed to predict the bounding box of the referred object. These methods suffer from the following two weaknesses: Firstly, the pre-trained visual feature extractors introduce text-unrelated visual signals into the visual features that hinder multimodal interaction. Secondly, the reasoning process followed in these two methods lacks visual guidance for language modeling. It is clear from these shortcomings that the reasoning ability of existing one-stage methods is limited. We propose a low-level interaction to extract text-related visual feature maps, and a high-level interaction to incorporate visual features in guiding the language modeling and further performing multistep reasoning on visual features. Based on the proposed interactions, we present a novel network architecture called the dual-path multilevel interaction network (DPMIN). Furthermore, experiments on five commonly used visual grounding datasets are conducted. The results demonstrate the superior performance of the proposed method and its real-time applicability.
The demands of deep neural network models for computation and storage make them unsuitable for deployment on embedded devices with limited area and power. To solve this issue, stochastic computing reduces the storage and computational complexity of neural networks by representing data as a stochastic sequence, followed by arithmetic operations such as addition and multiplication through basic logic operation units. However, short stochastic sequences may cause discretization errors when converting network weights from floating point numbers to the stochastic sequence, which can reduce the inference accuracy of stochastic computing network models. Longer stochastic sequences can improve the representation range of stochastic sequences and alleviate this problem, but they also result in longer computational latency and higher energy consumption. We propose a design for a differentiable quantization function based on the Fourier transform. The function improves the matching of the model to stochastic sequences during the network’s training process, reducing the discretization error during data conversion. This ensures the accuracy of stochastic computational neural networks with short stochastic sequences. Additionally, we present an adder designed to enhance the accuracy of the operation unit and parallelize computations by chunking inputs, thereby reducing latency. Experimental results demonstrate a 20% improvement in model inference accuracy compared to other methods, as well as a 50% reduction in computational latency.
The surface structure of a container lock pin is complex, making it difficult to establish a point cloud model with a high surface feature integrity. Therefore, a multi-view and multi-attitude point cloud model reconstruction algorithm based on a turntable was proposed to restore the complete surface features of the locking pin. Considering that sensors at a fixed height are paired with rotating turntables in most scenarios, the collected surface features are usually somewhat missing. Initially, the algorithm uses the parameter calibration results of the turntable to realize the multi-view three-dimensional point cloud stitching, and establishes a fixed attitude point cloud model. Then, through the proposed improved spherical projection algorithm, the positioning of the locking pin on the turntable is selected to establish a point cloud model under another posture. Finally, the point cloud model with multiple attitudes is integrated to improve its surface characteristics. Experimental results show that the proposed algorithm can build a lock-pin point cloud model with high surface feature integrity.
Infrared small-target detection has always been an important technology in infrared tracking systems. The current infrared approaches for small-target detection in complex backgrounds are prone to generating false alarms and exhibit sluggish detection speeds from the perspective of the human visual system. Using the multiscale local contrast measure using a local energy factor (MLCM-LEF) method, an infrared small-target detection method based on a double-layer local energy factor is proposed. The target detection was performed from the perspectives of the local energy difference and local brightness difference. The double-layer local energy factor was used to describe the difference between the small target and the background from the energy perspective, and the weighted luminance difference factor was used to detect the target from the brightness angle. The infrared small target was extracted by a two-dimensional Gaussian fusion of the processing results of the two approaches. Finally, the image mean and standard deviation were used for adaptive threshold segmentation to extract the small infrared target. In experimental tests on public datasets, this method improved the performance in suppressing background compared with the MLCM-LEF algorithm, DLEF (double-layer local energy factor) reduced the detection of a single frame time by one-third.
Three-dimensional point cloud semantic segmentation is an essential task for 3D visual perception and has been widely used in autonomous driving, augmented reality, and robotics. However, most methods work under a fully-supervised setting, which heavily relies on fully annotated datasets. Many weakly-supervised methods have utilized the pseudo-labeling method to retrain the model and reduce the labeling time consumption. However, the previous methods have failed to address the conformation bias induced by false pseudo labels. In this study, we proposed a novel weakly-supervised 3D point cloud semantic segmentation method based on group contrastive learning, constructing contrast between positive and negative sample groups selected from pseudo labels. The pseudo labels will compete with each other within the group contrastive learning, reducing the gradient contribution of falsely predicted pseudo labels. Results on three large-scale datasets show that our method outperforms state-of-the-art weakly-supervised methods with minimal labeling annotations and even surpasses the performance of some classic fully-supervised methods.
The remarkable achievements of deep learning in computer vision have led to significant development in example-based texture synthesis. The texture synthesis model using neural networks mainly includes local components, such as convolution and up/down sampling, which is unsuitable for capturing irregular structural attributes in non-stationary textures. Inspired by the frequency and space domain duality, a non-stationary texture synthesis method based on hidden layer Fourier convolution is proposed in this study. The proposed method uses the generative adversarial network as the basic architecture, performs feature splitting along the channel in the hidden layer, and builds a local branch in the image domain and a global branch in the frequency domain to consider visual perception and structural information. Experimental results show that this method can handle structurally challenging non-stationary texture exemplars. Compared with state-of-the-art methods, the method yielded better results in the learning and expansion of large-scale structures.
The image caption generation algorithm based on decoupling commonsense association aims to eliminate the interference of commonsense association between various types of entities on the model reasoning, and improve the fluency and accuracy of the generated description. Aiming at the relationship sentences in the current image description that conform to common sense but do not conform to the image content, the algorithm first uses a novel training method to improve the attention of the relationship detection model to the real relationship in the image and improve the accuracy of relationship reasoning. Then, a relation-aware entity interaction method was used to carry out targeted information interaction for entities with relationships, and the relationship information was strengthened. The experimental results show that the proposed algorithm can correct some commonsense false relationships, generate more accurate image captions, and obtain better experimental results on various evaluation indicators.
This paper presents a method for skinning in character animation, utilizing implicit surfaces, which is designed to deform animated models with skeleton and associated skinning weights．This method reconstructs the mesh around a given skeleton with the Hermite radial basis function and Poisson-disk sampling on surfaces．This process transforms the character’s volume into a set of localized 3D scalar fields and preserves the original mesh properties．Field functions are then constructed and employed to refine the results obtained from the geometric skinning technique．The implicit method, combined with two types of combination operators, generates realistic skin deformations around the human skeleton model finally．The method does not cause candy twist and joint swelling problems, and can handle skin collision and muscle protrusions．Due to its post-processing feature, this method is very suitable for animation generation in standard production pipeline．
Mainstream sentence classification algorithms rely on a single word vector model to obtain the feature vector representation of text, which leads to insufficient text mapping ability. Therefore, a multi-kernel learning method is used to fuse multiple text representations based on different word vectors to improve the accuracy of sentence classification. In the process of fusing different kernel functions, traditional kernel function coefficient optimization methods often lead to long training time and difficulty in finding a local optimum. To address this problem, a new kernel function coefficient optimization method that continuously approximates the optimal kernel function coefficient value based on parameter space segmentation and breadth first search was developed. In this study, a support vector machine (SVM) was used as a classifier to perform classification experiments on seven text datasets, and the experimental results showed that the multi-kernel learning classification results were significantly better than those of single-kernel learning. Moreover, the proposed optimization method performed better than traditional methods with less training cost.
We propose a neural network training framework called momentum-updated representation with reconstruction constraint for 3D (three-dimensional) object recognition using 2D (two-dimensional) images without angle labels. First, self-supervised learning is employed to address the lack of angle labels. Second, we use momentum updating based on a dynamic queue to maintain the stability of the object representation. Furthermore, the reconstruction constraint is applied to the learning process with an auto-encoder module, which enables the representation to capture more semantic information of the objects. Finally, during training, a dynamic queue reduction strategy is proposed for handling the imbalanced data distribution. Experiments on two popular multi-view datasets, ModelNet and ShapeNet, demonstrate that the proposed method outperforms existing methods.
The trajectory optimization of cycling is hindered by the errors of positioning equipment, riding habits of non-motor vehicles, and other factors. It leads to quality problems, such as abnormal data and missing positioning information in the riding trajectory, impacting the application of trajectory-based riding-map inference and riding-path planning. To solve these problems, this paper creates a framework for improving the quality of cycling-trajectory data, based on the construction of a grid index, screening of abnormal trajectory points, elimination of wandering trajectory segments, elimination of illegal trajectory segments, calibration of drift trajectory segments, and recovery of missing trajectory. Comparative and ablation experiments are conducted by using a real non-motor-vehicle cycling-trajectory dataset. The experimental results verify that the proposed method improves the accuracy of cycling-map inference.
Shared-nothing distributed databases are designed for the high scalability and high availability request of Internet-based applications. There have been significant achievements in shared-nothing distributed databases, but for some shared-nothing databases with stateless computation layers, long conflict-detection paths challenge database performance under high-contention workloads. To solve this problem, we design two methods, pre-lock and local cache, together with a high-contention detection module that allow high-contention to be quickly detected and the corresponding high-contention-handling strategy applied. Experiments show that our design and optimization for high-contention transaction-processing architecture can improve the performance of distributed databases under high-contention workloads.
The diagnostic method based on dual-view fundus imaging is widely used in diabetic retinopathy (DR) screening. This method effectively solves the problems of image occlusion and limited field of view under single-view. This paper proposes a learning method of feature fusion between dual-view images based on the attention mechanism to improve the accuracy of DR classification by effectively integrating different view information. Due to the small proportion of lesions in fundus images, the self-attention mechanism was introduced to enhance the learning of local lesion features. Moreover, a cross-attention mechanism is proposed to effectively utilize information between dual-view images to improve the classification of dual-view fundus images. Experiments were performed on the internal DFiD dataset and public DeepDRiD dataset. The proposed method can effectively improve the accuracy of DR classification and can be used for large-scale DR screening to assist doctors in achieving an efficient diagnosis.
The popularity of intelligent devices such as smartphones and surveillance cameras has led to serious face privacy problems. Face de-identification is considered an effective tool for protecting face privacy by concealing identity information. However, most de-identification methods lack explicit control and controllable changes in identifying de-identified face images, resulting in de-identified images that are inapplicable to face authentication and retrieval and other identity-related tasks. Therefore, this study proposes an identity inter-relationship-consistent face de-identification task in which the identity inter-relationship between two arbitrary de-identified faces maintained the same as before de-identification. To this end, a task-driven identity inter-relationship consistent generative adversarial network is introduced to generate de-identified faces with a consistent identity inter-relationship. A rotation-based de-identifier was designed to modify the original identity features to be de-identified with identity inter-relationship consistency. In addition, identity control loss is introduced to guarantee a precise identity generation using a de-identified generator. Qualitative and quantitative results show that our method achieves improvements compared with exiting methods for de-identifying de-identified faces as well as for maintaining their identity inter-relationship consistent.
Review text contains comprehensive user and item information and it has a great influence on users’ purchase decision. When users interact with different target items, they may show complex interests. Therefore, accurately extracting review semantic features and modeling the contextual interaction between items and users is critical for learning user preferences and item attributes. Focusing on enhancing the personalization capture and dynamic interest modeling abilities of recommender systems, and considering the usefulness of different features, we propose a hierarchical description-aware personalized recommendation (DAPR) algorithm. At the word level of review text, we design a personalized information selection network to extract important word semantic features. At the review level, we design a neural network based on a cross-attention mechanism to dynamically learn the usefulness of reviews, concatenate review summaries as descriptions, and devise a co-attention network to capture rich context-aware features. The analysis of five Amazon datasets reveal that the proposed method can achieve comparable recommendation performance to the baseline models.
Chinese sentiment analysis is one of important researches in natural language processing, which aims to discover the sentimental tendencies in the Chinese text. In recent years, research on Chinese text sentiment analysis has made great progress in efficiencies, but few studies have explored the characteristics of the language and downstream task requirements. Therefore, in view of the particularity of Chinese text and the requirements of sentiment analysis, using the Chinese text sentiment analysis method that integrates multi-granularity semantic features, such as characters, words, radicals, and part-of-speech is proposed. This introduces radical features and emotional part-of-speech features based on character and word features. Additionally, this integration uses bidirectional the long short-term memory network (BLSTM), attention mechanism and recurrent convolutional neural network (RCNN). The softmax function is used to predict the sentimental tendencies by integrating multi-granularity semantic features. The comparative experiment results on the NLPECC (natural language processing and Chinese computing) dataset showed that the F1 score of the proposed method was 84.80%, which improved the performance of the existing methods to some extent and completed the Chinese text sentiment analysis task.
Acute kidney injury is a clinical disease with a high morbidity rate, and early identification of potential patients can facilitate medical interventions to reduce morbidity and mortality. In recent years, electronic health records have been widely used to predict an individual’s potential risk. Most of the existing acute kidney injury prediction models tackle the issue of sparsity and irregularity in the physiological variables data by aggregating data or imputing the missing value, but ignore the patient’s health status implied by the missing information. Moreover, they do not consider the characteristics of and correlation between the various modalities. To solve the above issues, we present a multi-modal disease prediction model for acute kidney injury. The proposed model considers a variety of modal data, including physiological variables, disease, and demographic data. A new mask and time span based long short term memory (LSTM) network is designed to learn the time span and missing information of individual Physiological variables, and furthermore, to capture their numerical changes and frequency changes. The multi-head self-attention mechanism is introduced to promote interaction learning of each modality representation. Experiments on the real-world application of acute kidney injury risk prediction and mortality risk prediction demonstrate the effectiveness and rationality of the proposed model.
In real-world scenarios, various events in the news are not only too nuanced and complex to distinguish, but also involve multiple entities. To address these problems, previous event-centric methods are designed to detect events first and then extract arguments, relying on imperfect performance for event trigger detection; this process, however, is unfit to deal with the sheer volume of news in the real world. Given that the performance of named entity recognition (NER) is satisfactory, we shift our perspective from an event-centric to a target-centric view. This paper proposes a new task: target-dependent event detection (TDED), which aims to extract target entities and detect their corresponding events. We also propose a semantic and syntactic aware approach to support thousands of target entity extractions first and subsequently the detection of dozens of event types; this approach can be applied to data from massive corporations. Experimental results on a real-world Chinese financial dataset demonstrated that our model outperformed previous methods, particularly in complex scenarios.
中国综合性科技类核心期刊(北大核心)
