Table of Content

30 November 2021, Volume 39 Issue 6

Previous Issue    Next Issue

Intelligent Security Defense Theory and Technology in Special Region

Survey of Neural Network Watermarking

FENG Le, ZHU Renjie, WU Hanzhou, ZHANG Xinpeng, QIAN Zhenxing

2021, 39(6):  881-892.  doi:10.3969/j.issn.0255-8297.2021.06.001

Asbtract ( 2325 )   PDF (910KB) ( 431 )  

References | Related Articles | Metrics

This article sorts out the development context of neural network watermarking technology in recent years, and roughly classifies the mainstream methods into four categories, namely white box watermark, black box watermark, boxless watermark and fragile watermark. Specifically, this article reviews the evaluation indicators of neural network watermarking and these four different types of neural network watermarking technologies, discusses the advantages and disadvantages of existing neural network watermarking schemes, and looks forward to the future development trend.

Robust Coverless Information Hiding Based on Image Classification

DONG Tenglin, LI Xinran, YAO Heng, QIN Chuan

2021, 39(6):  893-905.  doi:10.3969/j.issn.0255-8297.2021.06.002

Asbtract ( 1817 )   PDF (10623KB) ( 285 )  

References | Related Articles | Metrics

Aiming at the problem that traditional information hiding is difficult to resist the detection of steganalysis algorithms, this paper proposes a coverless data hiding method based on image classification and scale-invariant feature transform (SIFT) extraction. Firstly, an original image database is classified by using Faster R-CNN to generate different kinds of sub-image databases. Secondly, a robust hashing scheme is designed by using the direction information of SIFT feature points of the image, with which the image hash value of each sub-image database is calculated, and all images in each sub-image databases are mapped to corresponding binary hash values. Finally, secret information is divided into several segments, and by comparing the secret information segments with the binary hash values of all the images, the images corresponding to the secret information segments are retrieved from the sub-image database. These images are transmitted to the receiver as the carrier containing the secret to complete the information hiding process. The receiver receives all the stego-images, and extracts the secret information according to the agreed hashing scheme. Experimental results and analysis show that the proposed method is robust to JPEG compression, Gaussian noise, salt and pepper noise, image scaling and other attacks, and the performance of hiding capacity is also improved.

Reversible Data Hiding in Encrypted Domain with Block Rearrangement

YANG Yang, LI Xinran, HU Jinchuan

2021, 39(6):  906-922.  doi:10.3969/j.issn.0255-8297.2021.06.003

Asbtract ( 1742 )   PDF (2010KB) ( 162 )  

References | Related Articles | Metrics

We propose a reversible data hiding scheme in encrypted domain based on block rearrangement in bit plane. Firstly, specific encryption methods, including scrambling and block rearranging, are applied to encrypt an original image. Secondly, scrambled bit planes are divided into non-overlapping blocks, including homogeneous blocks and nonhomogeneous blocks, and a label map is generated to identify these two kinds of blocks. Thirdly, after embedding necessary auxiliary data into embedded homogeneous blocks, data hider can embed additional data into the remaining available blocks. Based on the availability of encryption key and data hiding key, receivers can realize separable operations of data extraction efficiently and image recovery losslessly. Experimental results demonstrate that the proposed scheme gains improvement both in embedding capacity and in the quality of decrypted image.

Recent Advances in Text Steganography and Steganalysis

KANG Huixian, YI Biao, WU Hanzhou

2021, 39(6):  923-938.  doi:10.3969/j.issn.0255-8297.2021.06.004

Asbtract ( 3105 )   PDF (641KB) ( 828 )  

References | Related Articles | Metrics

This paper sorts out the development context of text steganography and steganalysis, and divides text steganography algorithms into two categories:modified text steganography and generative text steganography. The implementation process of the two types of algorithms is summarized, and the advantages and disadvantages of mainstream algorithms are analyzed from the aspects of rate-distortion performance and safety. Aiming at the two types of text steganography algorithms, the realization process of the corresponding steganalysis algorithms is summarized, and the development trend of text steganography and steganalysis is prospected.

Multi-target Detection and Recognition for Vehicle Inspection Images Based on Deep Learning

OU Qiaofeng, XIAO Jiabing, XIE Qunqun, XIONG Bangshu

2021, 39(6):  939-951.  doi:10.3969/j.issn.0255-8297.2021.06.005

Asbtract ( 1895 )   PDF (7360KB) ( 137 )  

References | Related Articles | Metrics

A multi-target detection and recognition method of vehicle inspection images based on deep learning is proposed for faster and more automatic vehicle inspection. Firstly, a lightweight yolov3 network is used to detect and recognize vehicle head, tires, license plate and triangle marks in a vehicle inspection image; secondly, a multi-task cascade convolution neural network is used to locate the four key points of the license plate; thirdly, according to the four key point coordinates and the size prior of the target license plate, the license plate image is corrected by perspective transformation; finally, a convolutional neural network is designed to classify the background color of the license plate. Thus, a convolutional recurrent neural network is realized for license plate character recognition. Experimental results show that the average end-to-end multi-target detection and recognition accuracy of this method is 98.03% on an 816×612 car inspection image. To facilitate the application of the deep learning model in vehicle inspection scenes, the model is deployed to a CPU using Alibaba reasoning engine, and the average speed of multi-target detection and recognition reaches 10 frames per second, which meets the application requirements of vehicle inspection.

Fundus Optic Disc Segmentation and Localization Based on Improved Multi-task Learning Method

LI Ning, SHANG Yingqiang, XIONG Jun, TAI Baoyu, SHI Chenjie

2021, 39(6):  952-960.  doi:10.3969/j.issn.0255-8297.2021.06.006

Asbtract ( 1908 )   PDF (3521KB) ( 92 )  

References | Related Articles | Metrics

In this paper, an improved multi-task learning method is proposed. The main structure of proposed network contains a feature extraction network and a dual path network for optic disc segmentation and location respectively. And through end-to-end training and testing, the multi-task purpose of automatic optic disc segmentation and location can be achieved. In the decoding phase of the feature extraction network, a dense layer is used to extract the context features of fundus images. The optic disc segmentation task relies on the decoding stage to gradually restore original image resolution and obtain the entire optic disc outline, whereas the optic disc localization task is to obtain accurate optic disc center coordinates by extracting context features of fundus images with atrous space pyramid module and pyramid pooling module. Optic disc segmentation and center localization of 350 fundus images are demonstrated. Experimental results show that the Dice coefficient between the automatically segmented optic disc results and manually marked optic disc areas is 0.965, and the average distance between the automatic localization and the manually marked optic disc center is 0.191 mm (34.7 pixels).

Signal and Information Processing

Progresses in Remote Sensing Information Extraction Methods for Rocky Desertification

CHONG Guoshuang, HAI Yue, ZHENG Hua, XU Weihua, OUYANG Zhiyun

2021, 39(6):  961-968.  doi:10.3969/j.issn.0255-8297.2021.06.007

Asbtract ( 1971 )   PDF (570KB) ( 311 )  

References | Related Articles | Metrics

In this paper, relevant literatures in the field of rocky desertification research are summarized, and methods of extracting rocky desertification information from remote sensing data are taken as the main line to analyze and discuss the evaluation index of rocky desertification and remote sensing information extraction of rocky desertification. This paper also summarizes the progress and shortcomings of human-computer interactive interpretation, supervised classification, unsupervised classification, object-oriented classification, and remote sensing data extraction of rocky desertification information, including multi-spectral remote sensing, hyperspectral remote sensing, microwave remote sensing and unmanned aerial vehicle remote sensing. The future development trend of remote sensing image information extraction methods in rocky desertification areas in my country is prospected. It is pointed out that more attention should be paid to the quantification of the classification index of rocky desertification degree, and the further research on the fusion of multi-source data features should be conducted to improve the accuracy of rocky desertification information extraction.

A Method for Aspect-Level Sentiment Classification Based on Multi-pattern Feature Fusion

FAN Shouxiang, YAO Junping, LI Xiaojun, CHENG Kaiyuan

2021, 39(6):  969-982.  doi:10.3969/j.issn.0255-8297.2021.06.008

Asbtract ( 1771 )   PDF (679KB) ( 178 )  

References | Related Articles | Metrics

Aiming at the problem of information loss in encoding long-distance texts as using recurrent neural networks, and the problem of attention bias to the high-frequency sentiment information when using the attention mechanism to extract sentiment information, this paper proposes a method of aspect-level sentiment classification based on multi-pattern feature fusion. The method divides sentiment information into three categories:single-point sentiment information, multi-point sentiment information and partial sentiment information with different expression patterns, and realizes mutual confirmation and error correction among various types of features by focusing on, extracting and fusing the three types of emotion information. The problems of information loss and attention bias are reduced, and the ability of aspect-level sentiment classification under complex sentiment expression patterns is enhanced. Experimental results show that the accuracy and F1 value of the aspect-level sentiment classification task in extracting and fusing sentiment information can be significantly improved by using the proposed method.

Research on Relaxed Polar Code Based on 3×3 High Dimensional Kernel Matrix

WEN Hao, CAO Yang

2021, 39(6):  983-994.  doi:10.3969/j.issn.0255-8297.2021.06.009

Asbtract ( 1706 )   PDF (695KB) ( 74 )  

References | Related Articles | Metrics

Aiming at the problem of high complexity in the polarization code constructed by the high-dimensional kernel matrix to improve the error correction performance, a construction scheme based on the 3×3 high-dimensional kernel matrix terminating the polarization code is proposed. Firstly, the kernel matrix G₅₃₁ with the highest polarization rate is selected to construct a termination polarization code, and it is theoretically proved on the binary erasure channel that the termination polarization code can reduce the computational complexity of encoding and decoding without affecting the error correction performance, and derive the upper and lower bounds of the complexity reduction ratio (CRC) of terminating polarization codes at the same time. Simulation shows that the complexity reduction ratio of the termination polarization code is related to the BEC (binary erasure channel) erasure probability. When the erasure probability is about 0.5, the complexity reduction ratio reaches the smallest, and the higher the target frame error rate (FER), the greater the complexity reduction ratio. When the FER is 10^-5, the highest complexity reduction rate of 71.43% can be achieved.

Control and System

Application of DDS Technology in Virtual Environment Chassis Dynamometer Test

ZHANG Xiaorui, ZHOU Zhili

2021, 39(6):  995-1005.  doi:10.3969/j.issn.0255-8297.2021.06.010

Asbtract ( 1717 )   PDF (3447KB) ( 87 )  

References | Related Articles | Metrics

There exist large amount of data and frequent transmission of data in virtual environment chassis dynamometer test system. To deal with the limitations of real-time data transmission and data throughput of the system, a data distribution service (DDS) technology is applied into the virtual environment chassis dynamometer test system. Based on the principle analysis of the virtual environment chassis dynamometer test and data transmission requirements, through the research of DDS technology and application, a virtual environment chassis dynamometer test system based on DDS technology is constructed. The key technical problems of DDS data transmission in the test system are solved through a series of implementations, including establishing subscribing models for different members, registering data types, defining themes, designing DDS interface classes, and encapsulating DDS interfaces. The data transmission delay and throughput of the test system are tested and analyzed, and test results show that the delay is only 9.4 ms and the data transmission throughput can reach 20 Mbit/s for the data volume reaches 4000 kB, which meets the requirements of vehicle chassis dynamometer virtual test system.

Sliding Mode Anti-swing Control for Unmanned Helicopter Slung-Load System Based on RBF Neural Networks

LIU Nan, CHEN Mou, WU Qingxian, SHAO Shuyi

2021, 39(6):  1006-1020.  doi:10.3969/j.issn.0255-8297.2021.06.011

Asbtract ( 1794 )   PDF (765KB) ( 324 )  

References | Related Articles | Metrics

Regarding unmanned helicopter slung-load systems which work in plane motion along with nonlinearity, strong coupling, unknown external bounded disturbance and modeling uncertainty, a sliding mode anti-swing control method based on radial basis function neural networks (RBFNNs) and a disturbance observer is proposed in this paper. Firstly, a system model is constructed in a general affine nonlinear form with its modeling uncertainty approximated by the RBFNNs. Secondly, the nonlinear disturbance observer is used to estimate the compound disturbance containing the approximation error of neural networks and external unknown bounded disturbance. Then a sliding mode anti-swing controller is designed based on RBFNNs and the disturbance observer. Furthermore, the stability of the closed-loop system is proved by using Lyapunov function. Finally, numerical simulations demonstrate the effectiveness of the control strategy.