T R A C K       P A P E R
Close

Login Panel

Close tab

Password Meter

Volume 11 Issue 02 (February 2024)

S.No. Title & Authors Page No View
1

Title : YOLO-SH:An Object Detection Model for Small Targets

Authors : Honggeng Zhang

Click Here For Abstract

Download Certificate
Abstract :

Achieving accurate object detection has always been a highly challenging task in the field of computer science, particularly in remote sensing image object detection. The possibility of objects appearing in any direction adds to the difficulty of achieving precise object detection. Additionally, the presence of small objects and the influence of environmental factors pose significant challenges for current deep learning-based object detection algorithms. In order to address these challenges, this paper proposes a target detection algorithm called YOLO-SH, which is based on YOLOv8. It improves the detection performance of objects appearing in any direction by incorporating the Swin Transformer structure into YOLO. The Swin Transformer utilizes sliding windows and hierarchical structures to achieve more efficient and flexible computation, and it can capture global information from multi-scale images with better feature extraction capabilities. Furthermore, to improve the detection performance for small objects, an additional prediction head is added, which effectively detects multi-scale objects based on the three original prediction heads in YOLOv8. Through extensive experiments on the DIOR dataset, our YOLO-SH model demonstrates excellent performance in remote sensing image detection.

1-5
2

Title : Social Media Rumor Detection Based on Multi-Channel Graph Convolutional

Authors : Meng LI, Mengyuan Liu, Lulu He

Click Here For Abstract

Download Certificate
Abstract :

The emergence of social media also speeds up the speed of information dissemination. At the same time, the development of social network platform also leads to the rapid spread of rumors, which has a great negative impact on social life. The research field of social media rumor detection has become a hot spot at present. The existing rumor detection methods can not effectively pay attention to the spread characteristics and breadth spread characteristics of social media rumors. To solve this problem, a new rumor detection method based on multi-channel graph Convolutional Network (MA-GCN) was proposed. The model mainly combines the textual features, propagation features and structural features of the event to judge whether the event is a rumor. The model explores the two characteristics of rumor propagation and dissemination through the operation of the top-down and bottom-up transmission modes of rumors, uses the bidirectional graph structure to learn the rumor propagation mode and capture the rumor propagation structure, uses the undirected graph to capture the global information for prediction, and introduces the multi-head attention mechanism for information fusion. The effectiveness of the proposed MA-GCN method is verified by experiments on twitter public datasets.

6-11
3

Title : IT-YOLOX: Object Detection Algorithm in UAV Perspective

Authors : Shuo Xu

Click Here For Abstract

Download Certificate
Abstract :

Recently, drone scenes have been widely used in rescue, agriculture and other industries, and drones can fly at different altitudes to obtain multiple fields of view and viewing angles. However, since most of the drones are shot at higher altitudes, there are many small targets and objects in the shooting images. In view of these difficulties, this paper proposes a new network model :IT-YOLOX, which is based on YOLOX-tiny. First, an additional detection layer is added to the original three detection layers of YOLOX-tiny to detect small target objects. Secondly, iAFF module is used in YOLOX-tiny neck to fuse the feature information of backbone network. Finally, transformer structure is used in the last part of the backbone network. This paper conducted experiments on VisDrone2019, and the experimental results proved that the improved network improved by 5.17% compared with the baseline model (YOLOX-tiny) mAP, especially strengthening the recognition ability of small targets.

12-16
4

Title : Stability Analysis for Discrete-Time Stochastic Neural Networks with time-varying delays

Authors : Xinran Ding

Click Here For Abstract

Download Certificate
Abstract :

This article investigates the stability analysis of discrete-time stochastic neural network systems with time-varying delays. In this neural network model, delays are time-varying. By establishing a new set of Lyapunov Krasovskii functionals and applying relevant lemmas, a criterion for robust global exponential stability related to delays in discrete-time stochastic neural network systems with time-varying delays is proposed, and presented in the form of linear matrix inequalities (LMIs), Transform the stability analysis problem to be solved into a feasibility problem for a set of linear matrix inequalities, and finally perform numerical validation using MATLAB to prove the effectiveness of the proposed method.

17-25
5

Title : BIM-Based Metro Station Pipeline Comprehensive Avoidance Optimization And Deepening Design Application Research

Authors : Bo Li, Zixun Qiao, Lan Xue

Click Here For Abstract

Download Certificate
Abstract :

As one of the important components of the metro project, the pipeline synthesis is characterized by a wide variety of specialties and intricate arrangement, and multiple units and specialties are involved in the design and construction of each in the design process, which leads to the situation that various types of information data are difficult to interact with each other. To address this problem, the pipeline installation of a metro station in Qingdao is taken as the research object, and BIM technology is applied to the design and deepening of pipeline synthesis in metro engineering to establish the algorithm of automatic pipeline collision avoidance optimization. We combine the HiBIM plug-in to carry on the collision detection to the established full-professional pipeline comprehensive BIM model of subway station to solve the problems such as errors, omissions and deficiencies in the comprehensive design of subway station pipelines. The results show that the developed pipeline avoidance algorithm solves a total of 442 collision problems and saves a lot of time for pipeline avoidance optimization; after the specific deepening design and application of the integrated model of all professional pipelines in metro stations, it brings significant results in improving the design efficiency and quality, and provides positive thinking for the integrated design and deepening of pipelines in metro stations.

26-32
6

Title : Research on Multi-modal Data Fusion Based on Swarm Learning

Authors : Tao Cheng, Haiyan Kang

Click Here For Abstract

Download Certificate
Abstract :

At present, the data of a single entity is not only increasing in data volume, but also in corresponding types of data modalities. In turn, applications involving multi-modal data fusion are increasing, and the corresponding privacy and security issues cannot be ignored. To this end, a Multi-modal Data Feature Fusion scheme based on Swarm Learning (MDFF-SL) is proposed to protect the privacy of multi-modal data. The SL-MDFF scheme puts all participants into a swarm learning architecture. First, the data features of two different modalities, namely text and image, are locally extracted, and then the respective data feature information is shared through the API interface in the swarm learning architecture for feature fusion. The entire solution ensures data privacy and security by virtue of the decentralization, tamper-proof, traceability and other characteristics of blockchain technology.

33-37
7

Title : Analysis of Chinese Vowel Pattern of Bangladeshi Students Based on Formants

Authors : Junying Niu

Click Here For Abstract

Download Certificate
Abstract :

With Chinas increasing international status, there has been a rise in the number of Bangladeshi students studying in China. Consequently, more students from Bangladesh are learning Chinese pronunciation to aid them in their work, studies, and daily life. Without proper instruction, they may encounter difficulties in learning Chinese pronunciation and make more errors. To enhance the Chinese language proficiency of Bangladeshi international students, this paper proposes a method for analysing the morphology of Chinese consonants based on their morphological features. The method analyses Chinese consonants from the perspective of their pronunciation, and the extracted consonant data are used to draw consonant coordinate maps and vowel space maps. Comparative analyses were conducted to identify differences in Chinese vowel pronunciation between Chinese and Bangladeshi students. The analysis provides a theoretical guide to correct pronunciation for Bangladeshi students.

38-41
8

Title : BMN_PAM: Boundary Matching Network with Pyramid Attention Module for Temporal Action Proposal Generation

Authors : Wu Han, Liang Jiayu

Click Here For Abstract

Download Certificate
Abstract :

Temporal action proposal generation (TAPG) is a task that aims to generate temporal action proposals, i.e. temporal segments that potentially contain actions, in untrimmed videos. It is crucial for various video analysis and understanding tasks, e.g. temporal action detection, video understanding. However, existing TAPG works generally fail to consider global dependencies of proposals and cannot capture multi-scale features of temporal actions. In this work, a new TAPG method is proposed, termed as BMN_PAM (Boundary Matching Network with Pyramid Attention Module), which can obtain multi-scale feature information and establish long-term/global dependencies between proposals. Specifically, BMN_PAM applies BMN as a baseline method to generate action boundary probabilities. In addition, a new PAM is designed to generate the confidence map of proposals, which exploits multi-scale features and global dependencies of proposals. Then, both the action boundary probabilities and the confidence map are combined to generate accurate action proposals. A benchmark TAPG dataset, i.e. ActivityNet-1.3, is used to evaluate the proposed method. Compared with five updated TAPG methods, BMN_PAM performs best with 75.72 in AR@100 (Average Recall) and 67.38 in AUC (Area Under Curve). In addition, BMN_PAM is generally better than BMN-based methods with other attention mechanisms.

42-48