Table of Content

30 May 2024, Volume 42 Issue 3

Previous Issue    Next Issue

Signal and Information Processing

Building Recognition of High-Resolution Remote Sensing Images Based on Deep Learning

LI Chengfan, MENG Lingkui, LIU Xuefeng

2024, 42(3):  375-387.  doi:10.3969/j.issn.0255-8297.2024.03.001

Asbtract ( 569 )   PDF (273KB) ( 381 )  

References | Related Articles | Metrics

This paper focuses on deep learning methods for detection and recognition of the buildings in the high-resolution remote sensing images. It provides a summary and analysis of the existing deep learning and building extraction methods and highlights future research directions. The study aims to contribute to the construction of sample libraries and remote sensing databases for deep learning-based target detection in high-resolution remote sensing images. This study can provide a reference for the construction of sample libraries and remote sensing databases for the deep learning-based target detection of highresolution remote sensing images. It also supports the building detection and recognition by deep learning in multi-scale and multi-source high-resolution remote sensing.

Estimating Flash Flood Disaster Susceptibility Based on K-means Clustering and Ensemble Learning Approaches

GUAN Zheng, YIN Yongqiang, ZHANG Xiaoxiang, CHEN Yuehong

2024, 42(3):  388-404.  doi:10.3969/j.issn.0255-8297.2024.03.002

Asbtract ( 266 )   PDF (10659KB) ( 106 )  

References | Related Articles | Metrics

In this paper, a model based on K-means clustering and ensemble learning approaches is developed to properly analyze the impact of spatial heterogeneity on the assessment of flash flood disaster susceptibility. Firstly, 12 338 catchments in Jiangxi Province, China, are selected as the study area, where the K-means clustering is performed on different frequency rainfall indicators for each period. Secondly, using the error sum of squares and mean contour coefficients as the clustering evaluation index, the small catchment datasets are divided into two subsets. Finally, for different subsets, ten flash flood influencing factors such as average slope, normalized difference vegetation index and rainfall are selected from geometric characteristics, environmental characteristics, and precipitation characteristics. The adaptive boosting (AdaBoost) and eXtreme gradient boosting (XGBoost) models are applied to evaluate the susceptibility of flash floods. It is found that precipitation is an important factor in flash floods disaster, and flash floods are more likely to occur in high precipitation areas in Jiangxi Province. Meanwhile, the distribution of high-risk areas is dispersed, mainly in the northeastern region and the northwestern edge. The area under the receiver operating characteristic curve (AUC) values of similar catchments could increase to 0.90 or above after clustering. The clustering model effectively addresses the heterogeneity of catchments as a precursor process for susceptibility assessment.

Vehicle-Mounted LiDAR Point Cloud Data Classification Based on Segmentation Algorithm and Graph Convolution Network

LIU Yawen, LIU Yongchang

2024, 42(3):  405-415.  doi:10.3969/j.issn.0255-8297.2024.03.003

Asbtract ( 179 )   PDF (1981KB) ( 67 )  

References | Related Articles | Metrics

Vehicle-mounted LiDAR point cloud semantic annotation is a prerequisite for further semantic analysis and understanding of road scenes. This paper proposes a point cloud classification method that integrates segmentation algorithm and graph convolution network. First, point cloud is segmented into point clusters using the density-based spatial clustering of applications with noise (DBSCAN) algorithm. Point clusters are treated as nodes, with adjacent clusters forming edges, thus constituting a graph. Then, graph convolution network is used to semi-supervise the classification of graph nodes and obtain the semantic annotation of any point in the point cloud. Experimental results demonstrate that replacing the original point cloud with point clusters greatly reduces the amount of data processed by the algorithm. Furthermore, the semi-supervised graph convolution network, considering the context of point cloud, achieves high classification accuracy even with a small number of labeled samples. The classification accuracy of experimental data with simple scenes is comparable to that of pointnet++, with the difference in accuracy below 6.7% for complex scenes.

Research on AR Tracking Method for Electronic Equipment Assembly Guidance

DU Xiaodong, WANG Peng, SHI Jiancheng, WANG Yue, SHUAI Hao

2024, 42(3):  416-424.  doi:10.3969/j.issn.0255-8297.2024.03.004

Asbtract ( 257 )   PDF (12103KB) ( 41 )  

References | Related Articles | Metrics

This paper aims to enhance the robustness and versatility of augmented reality (AR) tracking methods for electronic equipment assembly guidance by optimizing the structure of the position estimation network. This optimization involves integrating depthwise separable convolution with a channel attention mechanism. First, due to the lack of public datasets of 6 degrees of freedom (6-DOF) electronic equipment and various usage constraints, an RGB-D camera is used to collect and produce a 6-DOF training dataset for AR assembly guided electronic equipment. Then, using the structure of the position estimation network based on the pixel voting, depth-wise separable convolution is used to lighten the network, and the channel attention mechanism is introduced to evaluate the weight of the channels to compensate the accuracy loss caused by lightening the network. Finally, we verify the proposed network structure through AR assembly guidance by the electronic equipment task. Results show that the proposed tracking method exhibits superior robustness and maintains sound assembly guidance accuracy compared to existing method. Moreover, it can track the electronic equipment with weak texture and meet the real-time tracking requirements while ensuring accuracy.

5G Channel Propagation Model and Signal Coverage Inside Aircraft Cabin Scenarios

LIU Yuxin, YE Xijuan, BAO Junwei, MA Jian, CHEN Xiaomin, LI Mingsheng, ZHU Qiuming

2024, 42(3):  425-436.  doi:10.3969/j.issn.0255-8297.2024.03.005

Asbtract ( 225 )   PDF (7283KB) ( 54 )  

References | Related Articles | Metrics

In response to challenges encountered in wireless communication within aircraft cabins, including incomplete coverage, slow speeds, and instability, this study presents a 5th generation wireless communication technology (5G) channel model tailored for aircraft cabin scenarios using ray tracing methods. The signal coverage ability and channel parameter characteristics are analyzed. Firstly, we conduct three-dimensional geometric reconstruction of the real cabin scene using triangles to reduce the complexity of obtaining channel parameters by ray tracing. Subsequently, the 5G channel propagation model is constructed by combining the clustering algorithm, then the 5G signal coverage and communication performance inside aircraft cabin are analyzed. Simulation results show that the cluster power offset and the cluster time delay offset follow Gauss distribution, while the cluster azimuth angle of arrival offset and the cluster elevation angle of arrival offset follow Laplace distribution. Moreover, we found that the dense scatterers inside aircraft cabin are the key factors affecting 5G signal coverage. These conclusions can be used in the fields of radio signal coverage prediction and multipath parameter evaluation of 5G base station within aircraft cabin scenarios.

TEXAS Staging of Diabetic Foot Wounds Based on Deep Learning Approach

CHEN Yuqian, LYU Donghui, SONG Anping, XIE Chuantao

2024, 42(3):  437-446.  doi:10.3969/j.issn.0255-8297.2024.03.006

Asbtract ( 183 )   PDF (831KB) ( 206 )  

References | Related Articles | Metrics

In order to solve the problem of diabetic foot auxiliary diagnosis, an efficient deep learning method with two-level ensemble convolutional neural network was proposed. This paper proposes an efficient deep learning method featured with two-level ensemble convolutional neural networks. The approach utilizes DenseNet121 and EfficientNet-B0 networks with pre-training weight as initial parameters for feature extraction during network training. The Diabetic Foot Ulcers Grand Challenge 2021 dataset is used to train the parameters of whole network, so as to realize the automatic staging of diabetic foot in terms of wound infection and ischemia. 5-fold cross-validation was used to verify the proposed trained network. The proposed method achieves high accuracy, with AUC (area under the receiver operating characteristic curve) value, accuracy, recall, precision, and F1-score of the network measured as 0.989, 0.954, 0.944, 0.954, 0.956, respectively. The method demonstrates promising potential for assisting the staging of diabetic foot in clinical.

Analysis of Slope and Aspect with Elevation Anomaly in Guangdong Province

DENG Sisheng

2024, 42(3):  447-456.  doi:10.3969/j.issn.0255-8297.2024.03.007

Asbtract ( 171 )   PDF (1310KB) ( 40 )  

References | Related Articles | Metrics

The study introduces the application of digital elevation model (DEM) theory to analyze the morphological characteristics of quasi-geoids, focusing on slope, aspect, and other features. By establishing the relationship between elevation anomaly, horizontal distance, and azimuth, it delineates spatial variations in elevation anomalies, offering guidance for height measurement using global navigation satellite system (GNSS), digital height datum maintenance and other production practices. Based on DEM theory, slope and aspect of quasi geoid with different precisions and resolutions were extracted, reclassified, and statistical analyzed in Guangdong Province. Results indicate that quasi-geoid in Guangdong Province exhibits distinctive slope and aspect patterns, deviating significantly from standard DEMs. Overall, the value gradually decreases by 0.0020°, along an aspect of 297°, and its rate of change is approximately 3.49 cm/km. The slope and aspect of quasi geoid are strongly associated with its resolution and accuracy. The slope is more affected by resolution than precision, leading to reduced values as resolution coarsens.

Digital Media Forensics and Security

High-Performance Covert Communication Scheme Based on Image Cloud Data Deduplication

CHEN Yanghui, TANG Xin, ZHENG Tingting, CHANG Hanzhi, ZHOU Yiteng

2024, 42(3):  457-468.  doi:10.3969/j.issn.0255-8297.2024.03.008

Asbtract ( 190 )   PDF (1184KB) ( 52 )  

References | Related Articles | Metrics

This paper proposes a high-performance covert communication scheme based on image cloud data deduplication. Firstly, it uses images as the carrier of messages and proposes to map a file to more than one bit at a time based on the difference of the mean gray value of image regions, which enhances the covertness of the communication. Then, a set of basic image library based on grouping is designed, in which the images are arranged in a specified order. The dynamic interval determination mechanism is used to achieve the randomized selection of the message carrier, and the least significant bit algorithm is used to embed auxiliary marking information unrelated to the content of the message, ensuring communication security and reliability. Finally, a synchronization mechanism based on timestamp-aligned intervals is proposed. A multi-round traversal strategy and the concept of empty data grouping are introduced to effectively use the image library, which improve the transmission efficiency of communication. Comparative analysis with research utilizing cross-user deduplication for covert transmission reveals that the proposed scheme optimizes the correlation between the number of file uploads and the message length. Futhermore, it effectively improves the comprehensive transmission rate and enhances the security and the covertness of communication.

A Robust Coverless Image Steganography Method for Coding Camouflage

YUAN Ziye, QIU Baolin, YE Yu, WEN Wenying, HUA Dingli, ZHANG Yushu

2024, 42(3):  469-485.  doi:10.3969/j.issn.0255-8297.2024.03.009

Asbtract ( 261 )   PDF (4792KB) ( 47 )  

References | Related Articles | Metrics

Traditional image steganography methods are susceptible to attack by steganalysis tools, whereas coverless image steganography method can essentially resist the attack of steganalyzers. However, most coverless image steganography algorithms suffer from problems such as low robustness, limited extraction accuracy, and poor imperceptibility. Therefore, this paper proposes a robust coverless steganography method for coding camouflage, which combines depth-based synthetic steganography with traditional clustering algorithms. The proposed algorithm matches the synthetic images generated by the coding network with similar images through perceptual hashing, and converts the transmitted images from synthetic images to real natural images to improve security. In addition, clustering algorithm is used to find the camouflage image which is corresponding to the similar image for transmission. The clustering is based on the convolutional neural networks (CNN) feature, which improves the ability to resist geometric attacks. Experimental analysis demonstrates that the proposed scheme achieves higher capacity and extraction accuracy, and solves the problems of low image quality and poor robustness of generative steganography schemes.

A Novel Black-Box Finger-Print Watermarking Algorithm for Deep Classification Neural Network

MO Mouke, WANG Chuntao, GUO Qingwen, BIAN Shan

2024, 42(3):  486-498.  doi:10.3969/j.issn.0255-8297.2024.03.010

Asbtract ( 311 )   PDF (1896KB) ( 280 )  

References | Related Articles | Metrics

This paper proposes a novel framework and method for strong robust blackbox classification model finger-print watermarking. First of all, we develop a method for constructing poisoned images with high visual quality and enhanced security based on digital watermarking technology. This method embeds user identity information into the poisoned image, enabling traceability of deep neural network models in multiuser scenarios and reducing the susceptibility of the poisoned image to forgery. Second, we introduce a poisoned feature enhancement module to optimize the training of the model. Finally, we design an adversary training strategy, which can effectively learn the finger-print watermark with minimal embedding strength and reduce the probability of forged poisoned images. Extensive simulation experiments show that the good invisibility of the fingerprint watermark in the poisoned image constructed by our method, superior to similar optimal model watermarking methods such as WaNet. More than 99% of the black-box model finger-print watermarking verification rate is obtained at the cost of no more than a 2.4% reduction in the classification performance. Even with a difference of just one bit in the finger-print watermark, accurate verification of the model watermarking by copyright is achieved. These performances are generally better than the best-in-class model watermarking methods, demonstrating the feasibility and effectiveness of our proposed method.

Computer Science and Applications

A Multi-objective Flow QoS Scheduling Strategy with Improved Proximal Optimization

LIU Xingtong, ZHENG Hong, HUANG Jianhua

2024, 42(3):  499-512.  doi:10.3969/j.issn.0255-8297.2024.03.011

Asbtract ( 211 )   PDF (1786KB) ( 92 )  

References | Related Articles | Metrics

Software-defined networking (SDN) can be equipped with flexible flow scheduling strategies to improve the quality of network service systems. However, as the complexity of business traffic increases, existing flow scheduling algorithms may suffer from performance degradation due to decreased scene matching. To address this problem, this paper proposes an intelligent routing strategy based on deep reinforcement learning. The strategy collects various link information through SDN, and implements feature extraction and state awareness based on long-short term memory networks and proximal policy optimization algorithms. The strategy generates dynamic flow scheduling strategies that meet quality of service (QoS) goals in business scenarios, thereby maximizing QoS. Experimental results show that the proposed scheme enhances the QoS index of the entire system by 7.06% compared to existing routing strategies, effectively improving the throughput of the business system.

xDeepFM Recommendation Model Based on Field Factorization

LI Zijie, ZHANG Shu, OUYANG Zhaoxiang, WANG Jun, WU Di

2024, 42(3):  513-524.  doi:10.3969/j.issn.0255-8297.2024.03.012

Asbtract ( 195 )   PDF (711KB) ( 61 )  

References | Related Articles | Metrics

The eXtreme deep factorization machine (xDeepFM) is a context-aware recommendation model integrating a compressed interaction network for controllable feature cross-ordering. The network is combined with deep neural network to optimize the recommendation performance. To further improve xDeepFM’s performance in recommended scenarios, eXtreme deep field factorization machine (xDeepFFM) is proposed in this paper. The improved model enhances feature expression capabilities through field information and uses multiple compressed interaction networks to learn higher-order combinatorial features based on field information. Furthermore, this paper analyzes the rationality of the setting of user field and item field. The effectiveness of the improved model is evaluated using area under curve and Log-likelihood loss metrics on three public datasets of different sizes.

Recommendation Algorithm Based on Fuzzy Preference Label Vector

SU Zhan, YANG Haochuan, AI Jun

2024, 42(3):  525-539.  doi:10.3969/j.issn.0255-8297.2024.03.013

Asbtract ( 146 )   PDF (1079KB) ( 34 )  

References | Related Articles | Metrics

Traditional collaborative filtering algorithms suffer from various shortcomings such as insufficient accuracy due to rating prediction errors, and limited scalability of the algorithm due to the need to cache numerous similarity results. To address these challenges, this paper proposes a user fuzzy preference similarity measurement method based on item label vectors. The method uses fuzzy logic to measure the likes and dislikes of different users towards different item content labels, and represents the similarity between users as a vector across these labels. The corresponding similarity calculation and rating prediction formulas are then designed based on the relationship between this vector and the predicted target item content labels. Experimental results on two commonly used datasets demonstrate significant improvements over recent algorithms. Specifically, compared to existing methods, the proposed approach yields a 12.38% enhancement in mean absolute error, indicating improved rating prediction accuracy, a 7.85% increase in F1 value, indicating enhanced preference prediction accuracy, and a 17.47% improvement in half-life utility, reflecting sorting accuracy. Furthermore, the algorithm proposed in this paper reduces the number of neighbors required for the optimal prediction of each metric, thereby shortening the running time of the algorithm and effectively enhancing the scalability.

Traffic Flow Prediction Method of Highway Toll Station Based on GAT-LSTM Model

LIU Shengqing, MA Feihu

2024, 42(3):  540-548.  doi:10.3969/j.issn.0255-8297.2024.03.014

Asbtract ( 495 )   PDF (225KB) ( 184 )  

References | Related Articles | Metrics

To achieve accurate prediction of highway station traffic flow, this paper proposes a research method that utilizes a combined model to capture the spatio-temporal characteristics of highway toll station traffic. The basic idea is to mine the toll data to obtain a spatio-temporal dataset of traffic flow, analyze its spatio-temporal characteristics, reveal the spatio-temporal evolution rules and correlation mechanisms between the traffic flow of highway toll stations. Subsequently, we combine these insights with deep learning models to predict highway traffic flow. As a case study, we focus on the main toll station in Jiujiang, Jiangxi Province, utilizing the toll data from May 1, 2021, to December 31, 2021. The extracted spatio-temporal traffic data serves as input for our model, yielding analysis and prediction results of outlet flow. The prediction performance of the model is evaluated through three indicators: mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE). The results show that the proposed model effectively improves the prediction accuracy by utilizing spatio-temporal characteristics, outperforming single models in predictive capability.