Table of Content

30 January 2024, Volume 42 Issue 1

Previous Issue    Next Issue

Special Issue on Computer Application

Object Detection Based on Nonlinear Gaussian Squared Distance Loss

LI Rui, LI Yi

2024, 42(1):  1-14.  doi:10.3969/j.issn.0255-8297.2024.01.001

Asbtract ( 77 )   PDF (7094KB) ( 61 )  

References | Related Articles | Metrics

Existing series of loss functions based on intersection over union (IoU) have certain limitations, impacting the accuracy and stability of bounding box regression in object detection. To address this problem, a bounding box regression loss based on nonlinear Gaussian squared distance is proposed. Firstly, the three factors including overlapping, center point distance and aspect ratio in the bounding box are comprehensively considered, and the bounding box is modeled as a Gaussian distribution. Then a Gaussian squared distance is proposed to measure the distance between two distributions. Finally, a nonlinear function is designed to transform the Gaussian square distance into a loss function that facilitates neural network learning. Experimental results show that compared with IoU loss, the mean average precision of the proposed method on mask region-based convolutional neural network, fully convolutional one-stage object detector and adaptive training sample selection object detector is improved by 0.3%, 1.1% and 2.3%, respectively. These results demonstrate the efficiency of the proposed method in enhancing target detection performance and supporting the regression of high-precision bounding boxes.

An Automatic Atrial Fibrillation Detection Model Based on GAN and MS-ResNet

QIN Jing, HAN Yue, WANG Liyong, JI Changqing, LIU Lu, WANG Zumin

2024, 42(1):  15-26.  doi:10.3969/j.issn.0255-8297.2024.01.002

Asbtract ( 85 )   PDF (3052KB) ( 70 )  

References | Related Articles | Metrics

Atrial fibrillation (AF) is a common cardiac arrhythmia. However, existing research often relies on single-scale signal segments and overlooks potential complementary information at different scales as well as data imbalance issues, leading to decreased diagnostic performance. This paper proposes a novel AF automatic detection model based on generative adversarial network (GAN) and residual multi-scale network. The model utilizes GAN to synthesize single-lead ECG data with high morphological similarity, hence addressing data privacy and imbalance issues. A multi-scale residual network (MS-ResNet) feature extraction strategy was designed to extract the features of signal segments of different sizes from various scales, so as to effectively capture the features of P wave disappearance and RR interval irregularity. The model combines these two strategies not only to generate high-quality ECG (electrocardiogram) data for the automatic AF detection but also to extract temporal features between different waves using multi-scale grids. The performance of MS-ResNet is evaluated on the PhysioNet Challenge 2017 public ECG dataset and a balanced dataset, comparing it with other existing atrial fibrillation classification models. Experimental results show that the average F1 value and accuracy rate of MS-ResNet on the balanced dataset are 0.914 1 and 91.56%, respectively. Compared with the unbalanced dataset, F1 increases by 4.5%, and the accuracy rate increases by 3.5%.

Semi-supervised Rock Slice Image Classification Based on Hierarchy Consistency Mean Teacher Model

YAN Zijie, WANG Yang, CHEN Yan, ZHANG Chong

2024, 42(1):  27-38.  doi:10.3969/j.issn.0255-8297.2024.01.003

Asbtract ( 75 )   PDF (10754KB) ( 46 )  

References | Related Articles | Metrics

Traditional rock slice image classification relies on a large number of manually labeled image samples, which is subject to the experience and ability of the labelers. This practice limits the scalability of classification enhancement as increasing unlabeled rock slice image samples does not contribute effectively. In order to achieve effective utilization of unlabeled data information, the hierarchy consistency mean teacher (HCMT) model adds a hierarchy consistency regularization term to the unsupervised loss of the mean teacher (MT) model to constrain the hierarchical structure of the teacher-student model. Ablation experiments and comparative analyses reveal that the introduction of hierarchy consistency regularization method improves the classification performance of the MT model by using the effective information of unlabeled data. As a result, the HCMT model achieves comparable classification capability in both half-labeled and fully labeled dataset. The experiments show the potential of the semi-supervised learning model to improve the classification ability of the model by using a large number of unlabeled rock slice image data.

Quantum Attacks on Symmetric Cryptosystems

FENG Xiaoning, WU Hongyu

2024, 42(1):  39-52.  doi:10.3969/j.issn.0255-8297.2024.01.004

Asbtract ( 85 )   PDF (682KB) ( 44 )  

References | Related Articles | Metrics

This paper undertakes an investigation of recent research trends in quantum attacks on symmetric encryption schemes, offering an analysis of the connections between mainstream attack methods and various literature sources. Mainstream attack methods are systematically categorized into three types: quantum period attacks, Grover algorithmrelated attacks, and quantum differential attacks. For each category, representative attack methods are introduced, accompanied by an elucidation of the core concepts underlying each approach. Furthermore, we contemplate future research directions within this domain, considering potential advancements in light of existing attack schemes.

Server Energy Consumption Model Based on ConvLSTM in Mobile Edge Computing

LI Xiaolong, LI Xi, YANG Lingfeng, HUANG Hua

2024, 42(1):  53-66.  doi:10.3969/j.issn.0255-8297.2024.01.005

Asbtract ( 68 )   PDF (974KB) ( 48 )  

References | Related Articles | Metrics

To address the issue of low sensitivity and accuracy of existing energy consumption models in accommodating dynamic workload fluctuations, this paper proposes an intelligence server energy consumption model (IECM) based on the convolutional long short-term memory (ConvLSTM) neural network in mobile edge computing, which is used to predict and optimize energy consumption in servers. By collecting server runtime parameters and using the entropy method to filter and retain parameters significantly affecting server energy consumption, a deep network for training the server energy consumption model is constructed based on the selected parameters using the ConvLSTM neural network. Compared with existing energy consumption models, IECM exhibits superior adaptability to dynamic changes in server workload in CPU-intensive, I/O-intensive, memoryintensive, and mixed tasks, offering enhanced accuracy in energy consumption prediction.

Target Counting Method Based on UAV View in Large Area Scenes

XIE Ting, ZHANG Shoulong, DING Laihui, XU Zhiwei, YANG Xiaogang, WANG Shengke

2024, 42(1):  67-82.  doi:10.3969/j.issn.0255-8297.2024.01.006

Asbtract ( 84 )   PDF (9343KB) ( 79 )  

References | Related Articles | Metrics

In recent years, unmanned aerial vehicles (UAVs) have been widely used in the field of crowd counting due to their high flexibility and maneuverability. However, most of the existing crowd counting methods are based on single viewpoints, with limited studies focusing on multi-viewpoint counting in large-scale, multi-camera scenes. To solve this problem, this paper proposes a UAV-based target counting method which can accurately count the number of targets in a given scene. Specifically, this study selects a sea-front area for data acquisition, utilizes deep learning technology for target detection and image stitching fusion on the acquired images. The detection information is then mapped onto the spliced image, and a counting algorithm is employed to fulfill the counting task for the regional scene. The effectiveness of the counting algorithm based on target detection is validated through experiments conducted on both public dataset and the dataset produced in this paper.

Research on Enhanced Routing for Reinforcement Learning in Wireless Sensor Networks

ZHANG Huanan, LI Shijun, JIN Hong

2024, 42(1):  83-93.  doi:10.3969/j.issn.0255-8297.2024.01.007

Asbtract ( 56 )   PDF (648KB) ( 45 )  

References | Related Articles | Metrics

The classical problem of finding the optimal parent node in wireless network tree routing is discussed in this study. Various indexes affecting the decision rules of tree routing are analyzed, such as weighted average received signal strength, buffer occupation rate and power consumption ratio. A system model of enhanced tree routing protocol and reinforcement learning algorithm based on reinforcement learning is proposed in wireless sensor networks. The basic operation of the proposed tree-based routing protocol is described in detail, and the algorithm is updated for cyclic detection of parent node. In order to make adaptive decisions in complex scenarios, a state space, an action set and an excitation function are defined. The optimal parent node with the highest excitation is identified through trial and error. Through simulation and comparative study, it is verified that the parent node selection scheme achieves reasonable tradeoff among the performance indicators such as end-to-end delay, reliability and energy consumption. Through simulation and comparative analysis, the efficacy of the parent node selection scheme is validated, demonstrating a judicious tradeoff among performance indicators such as end-to-end delay, reliability, and energy consumption.

Research on Different Desensitization Data Based on Federated Ensemble Algorithm

LUO Changyin, CHEN Xuebin, ZHANG Shufen, YIN Zhiqiang, SHI Yi, LI Fengjun

2024, 42(1):  94-102.  doi:10.3969/j.issn.0255-8297.2024.01.008

Asbtract ( 61 )   PDF (542KB) ( 49 )  

References | Related Articles | Metrics

To solve the problem that gradient updating leads to the possible leakage of local data in federated learning, federated ensemble algorithms based on local desensitization data are proposed. The algorithm desensitizes the raw data with different values of variability and fitness thresholds, employing diverse models for local training on data with different desensitization levels to ascertain parameters suitable for a federated ensemble approach. Experimental results show that the stacking federated ensemble algorithm and voting federated integration algorithm outperform the baseline accuracy achieved by the federated average algorithm with traditional centralized training. In practical applications, different desensitization parameters can be set according to different needs to protect data and improve its security.

Object Tracking Algorithm Based on Vehicle Appearance Features and Inter-frame Optical Flow

LI Shaoqian, CHENG Xin, ZHOU Jingmei, ZHAO Xiangmo

2024, 42(1):  103-118.  doi:10.3969/j.issn.0255-8297.2024.01.009

Asbtract ( 57 )   PDF (70923KB) ( 36 )  

References | Related Articles | Metrics

In complex road scenes, frequent occlusions and similar appearances between vehicle targets, coupled with the use of static preset parameters used throughout the entire movement of the targets collectively contribute to a decline in tracking accuracy. This paper proposes an object tracking algorithm based on vehicle appearance features and inter-frame optical flow. Firstly, the position information of the vehicle target frame is obtained through the YOLOv5x network model. Secondly, the optical flow between the current frame and the previous frame is calculated using the RAFT (recurrent all-pairs field transforms for optical flow) algorithm, and the optical flow map is clipped according to the obtained position information. Finally, in the process of Kalman filtering, inter-frame optical flow is used to compensate for more accurate motion state information, while vehicle appearance features and intersection over union (IOU) features are used to complete trajectory matching. Experimental results show that the tracking algorithm correlating inter-frame optical flow performs well on the MOT16 data set. Compared with simple online and realtime tracking with a deep association metric (DeepSORT), mostly tracked trajectories (MT) has increased by 1.6%, multiple object tracking accuracy (MOTA) has increased by 1.3%, and multiple object tracking precision (MOTP) has increased by 0.6%. The accuracy of the improved vehicle appearance feature extraction model has been improved by 1.7% and 6.3% on the training and verification sets, respectively. Consequently, leveraging the high-precision vehicle appearance feature model and motion state information from the associated inter-frame optical flow enables effective vehicle target tracking in traffic scenes.

Knowledge Graph Completion Method Based on Semantic Hierarchy in Spherical Coordinates

GUO Ziyi, ZHU Tong, LIN Guangyan, TAN Huobin

2024, 42(1):  119-133.  doi:10.3969/j.issn.0255-8297.2024.01.010

Asbtract ( 54 )   PDF (1249KB) ( 74 )  

References | Related Articles | Metrics

Most of existing knowledge graph completion methods often neglect the semantic hierarchical differences that objectively exist between entities. To address these limitations, we propose a knowledge graph completion method named spherical hierarchical knowledge completion (SpHKC), which models semantic hierarchical phenomena using spherical coordinates. In this method, entities are mapped to points on a spherical surface, and entities located on different spheres correspond to different semantic hierarchy levels. The radius of the sphere determines the level of the semantic hierarchy for entities on that sphere, with larger spheres representing lower levels. Relationships are modeled as movements from one entity on the spherical surface to another entity (on the same or different spheres), involving rotation and positioning operations to handle both similar and different semantic hierarchy levels between entities. The polar angle and azimuth angle in spherical coordinates provide entities with richer expressions. Experimental results demonstrate that SpHKC achieves comparable performance to state-of-the-art methods on the FB15k-237 and WN18RR datasets. Moreover, it consistently improves important metrics such as MRR (mean reciprocal ranking) and Hits@10 by approximately 1% compared to recent algorithms on the YAGO3-10 dataset, showcasing the effectiveness of incorporating semantic hierarchical information.

Browser Power Optimization Based on CPU-GPU Co-regulation and Web Page Feature Perception

ZHANG Jin, HUANG Jiangjie, PENG Long, LIU Xiaodong, YU Jie, HUANG Haowei, WANG Wenzhu

2024, 42(1):  134-144.  doi:10.3969/j.issn.0255-8297.2024.01.011

Asbtract ( 72 )   PDF (755KB) ( 38 )  

References | Related Articles | Metrics

Android's inability to sense web page content during resources allocation to the browser often results in over-allocation of resources and unnecessary loss of power. At the same time, due to the growth of CPU adjustable frequency density, optimizing energy consumption through dynamic voltage and frequency scaling (DVFS) technology becomes increasingly challenging. Furthermore, the role of the graphics processing unit (GPU) in browser operation is ignored under the system's default regulation policy. Aiming at the above problems, we propose a method to optimize power consumption by co-regulating CPU and GPU. First, web pages are classified by logistic regression based on the processor operating characteristics when loading web pages. We assign weights to webpage characteristics to quantify the complexity, and then use DVFS to limit the CPU frequency while adjusting the GPU frequency based on webpage category and complexity. The proposed method is applied to the Chromium browser on Google Pixel2 XL, and tested on the top 500 Chinese websites, resulting in a 12% reduction in power consumption and an average 5% decrease in webpage loading time.

Track Area Detection for Railway Switches

CHEN Yijun, CHEN Yu, TENG Fei

2024, 42(1):  145-160.  doi:10.3969/j.issn.0255-8297.2024.01.012

Asbtract ( 56 )   PDF (12570KB) ( 54 )  

References | Related Articles | Metrics

The detection of the railway track area in front of the train is a key link in active train collision avoidance technology. The existing railway area segmentation methods are mostly used for track detection in simple scenarios, posing challenges when confronted with complex scenarios such as railway switches in actual operation. We propose a method for detecting railway track areas for railway switches, which solves the problem that existing technology encounters difficulty in detecting the actual running area of trains under railway switches. First, a railway track area segmentation model based on information fusion is proposed. Aiming at the problem of difficulty in matching the left and right rails of the railway, the railway area and the rails are segmented and the segmentation results are used for rail matching. Second, a railway area reconstruction method based on inverse perspective transformation is designed to reconstruct the railway area by preserving the key points of the rails. Meanwhile, a railway switch classification model based on grouped convolution is used to identify the switch direction. Experimental results show that the proposed method achieves high accuracy in complex environments, with pixel accuracy (PA) index of 98.67% and Mean Intersection over Union (MioU) index of 98.12%, showcasing its potential applicability to trains.

A Multi-label Semantic Calibration Method for Few Shot Extractive Question

LIU Qing, CHEN Yanping, ZOU Anqi, QIN Yongbin, HUANG Ruizhang

2024, 42(1):  161-173.  doi:10.3969/j.issn.0255-8297.2024.01.013

Asbtract ( 56 )   PDF (1295KB) ( 24 )  

References | Related Articles | Metrics

biases, especially in instances involving multiple sets of distinct repeated spans. Therefore, this paper proposes a multi-label semantic calibration method for few-shot extractive QA to mitigate the above issues. Specifically, this method uses the head label, which contains global semantic information, and the special character in the baseline model to form a multi-label for semantic fusion. The semantic fusion gate is then used to control the introduction of global information flow to integrate global semantic information into the semantic information of the special character. Next, the semantic selection gate is used to retain or replace the newly integrated global semantic information and the original semantic information of the special character, achieving semantic adjustment of label bias. The results of 56 experiments on 8 few-shot extractive QA datasets consistently outperformed the baseline model in terms of the evaluation metric F1 score. This demonstrates the effectiveness and advancement of the method.

Projected Reward for Multi-robot Formation and Obstacle Avoidance

GE Xing, QIN Li, SHA Ying

2024, 42(1):  174-188.  doi:10.3969/j.issn.0255-8297.2024.01.014

Asbtract ( 72 )   PDF (1502KB) ( 78 )  

References | Related Articles | Metrics

To address issues of excessive centralization, low system robustness, and formation instability in multi-robot formation tasks, this paper introduces the projected reward for multi-robot formation and obstacle avoidance (PRMFO) approach. PRMFO achieves decentralized decision-making for multi-robot using a unified state representation method, ensuring consistency in processing information regarding interactions between robots and the external environment. The projected reward mechanism, based on this unified state representation, enhances the decision-making foundation by vectorizing rewards in both distance and direction dimensions. To mitigate excessive centralization, an autonomous decision layer is established by integrating the soft actor-critic (SAC) algorithm with uniform state representation and the projected reward mechanism. Simulation results in the robot operating system (ROS) environment demonstrate that PRMFO enhances average return, success rate, and time metrics by 42%, 8%, and 9%, respectively. Moreover, PRMFO keeps the multi-robot formation error within the range of 0 to 0.06, achieving a high level of accuracy.