INTERFERENCE MITIGATING BASED ON FRACTIONAL FOURIER TRANSFORM IN TRANSFORM DOMAIN COMMUNICATION SYSTEM

The method of FRactional Fourier Transform (FRFT) is introduced to Transform Domain Communication System (TDCS) for signal transforming in the paper after theoretical analysis. The method yields optimal Basis Function (BF) by FRFT with optimal transform angle. The TDCS using the proposed method has wider usable spectrum, stronger robustness and better ability of anti non-stationary jamming than using usual methods, such as Fourier Transform (FT), Auto Regressive (AR), Wavelet Transform (WT), etc. The main simulation results are as follows. First, the Bit Error Rate (BER) Pb is close to theoretical bound of no jamming no matter in single tone or in linear chirp interference. Second, the interference-to-signal ratio J /E is at least 12dB more than that of Direct Spread Spectrum System (DSSS) under the same BER if the spectrum hopping-to-signal ratio is 1:20 in chirp plus hopping interfering. Third, the Eb /N 0(when estimation difference is 90% between trans- mitter and receiver) is about 3.5dB or about 0.5dB (when estimation difference is 10% between transmitter and receiver) more than that of theoretical result when no estimation difference un-der Pb=10-2.

Speech Encryption in Linear Canonical Transform Domain Based on Chaotic Dynamic Modulation

In order to transmit the speech information safely in the channel,a new speech encryp-tion algorithm in linear canonical transform(LCT)domain based on dynamic modulation of chaot-ic system is proposed.The algorithm first uses a chaotic system to obtain the number of sampling points of the grouped encrypted signal.Then three chaotic systems are used to modulate the corres-ponding parameters of the LCT,and each group of transform parameters corresponds to a group of encrypted signals.Thus,each group of signals is transformed by LCT with different parameters.Fi-nally,chaotic encryption is performed on the LCT domain spectrum of each group of signals,to realize the overall encryption of the speech signal.The experimental results show that the proposed algorithm is extremely sensitive to the keys and has a larger key space.Compared with the original signal,the waveform and LCT domain spectrum of obtained encrypted signal are distributed more uniformly and have less correlation,which can realize the safe transmission of speech signals.

TRANSFORM DOMAIN SMART ANTENNAS ALGORITHM FOR MAI SUPPRESSION

Multiple Access Interference(MAI) is the major factor that degrades the performance of a CDMA system. In this paper, a novel transform domain algorithm combined with parameter estimation for MAI suppression is proposed. Compared with the method that combines an adaptive array antenna with parameter estimation for interference suppression, it converges faster with the same Bit Error Rate(BER) performance.

LAPPED TRANSFORM DOMAIN INTERFERENCE EXCISION USING MODIFIED MEDIAN FILTER

A novel communication receiver which uses lapped transform(LT) incorporating modified median filter(MMF) algorithm was designed for narrow band interference(NBI) excision.Comparing to traditional Fourier Transform,LT has longer basis vectors,less spectral leakage,thus better frequency resolution.The LT domain MMF algorithm takes full advantages of the direct sequence spread spectrum signal,as well as the characteristics of LT,performs the transform domain filtering twice.The first filtering locates the position of interference and mitigates most of them.The second filtering is performed in a small neighborhood of the located interference.So LT domain MMF algorithm can completely mitigate the interference without distorting the desired signal.The simulation results demonstrate the improved BER(Bit Error Rate)performance and increased robustness of our receiver.

基于Transformer-GRU并行网络的滚动轴承剩余寿命预测

为有效描述滚动轴承性能退化趋势和准确预测其剩余寿命,提出一种基于多域特征融合的Transformer-GRU并行网络的滚动轴承剩余寿命预测方法。建立评价指标对滚动轴承振动信号的时域、频域和时频域等多域特征进行筛选,得到评分高的敏感特征,获得退化特征集。利用自编码对退化特征集进行降维,减少数据复杂度和冗余度,得到滚动轴承的退化曲线。最后,利用Transformer-GRU并行网络进行剩余寿命预测,并将该方法运用到公开的轴承数据集分析中。结果表明:Transformer-GRU并行网络不仅可以高效准确地捕捉输入序列中的长期依赖关系,还能更好地处理时间序列之间的特征;该方法可以有效地预测滚动轴承剩余寿命,相比LSTM、GRU等经典方法更具优越性和泛化性。

基于双域Transformer耦合特征学习的CT截断数据重建模型

目的为解决CT扫描视野(FOV)不足导致的截断伪影和图像结构失真问题,本文提出了一种基于投影和图像双域Transformer耦合特征学习的CT截断数据重建模型(DDTrans)。方法基于Transformer网络分别构建投影域和图像域恢复模型,利用Transformer注意力模块的远距离依赖建模能力捕捉全局结构特征来恢复投影数据信息,增强重建图像。在投影域和图像域网络之间构建可微Radon反投影算子层,使得DDTrans能够进行端到端训练。此外,引入投影一致性损失来约束图像前投影结果,进一步提升图像重建的准确性。结果Mayo仿真数据实验结果表明,在部分截断和内扫描两种截断情况下,本文方法DDTrans在去除FOV边缘的截断伪影和恢复FOV外部信息等方面效果均优于对比算法。结论DDTrans模型可以有效去除CT截断伪影,确保FOV内数据的精确重建,同时实现FOV外部数据的近似重建。

基于双域交互Transformer的磁共振图像重建

对k空间数据部分采样是加速磁共振成像的主要方法.从欠采样的数据中重建出高质量的磁共振图像,在临床诊断和研究分析中有着重要的应用价值.近年来,基于深度学习的方法在磁共振重建领域取得了一些进展,然而单独面向图像域或频域的网络不能同时利用双域特征共同提升重建质量.另外,虽然已有一些双域重建方法模型,但是缺乏双域数据交互融合限制了重建性能.针对以上问题,本文提出了一种基于双域交互Transformer的磁共振图像重建网络模型,使用Transformer提取双域特征,并利用交互注意力引导双域特征融合,实现了双域特征的高效提取和交互.具体地,首先,由于频域数据每个点都对应着图像域的所有像素点,因此在k空间使用1×1的卷积提取全局特征,同时使用基于窗口的Transformer将注意力的计算限制在了窗口中,减小了计算负担,并且可以有效地对图像域特征进行表示.其次,提出了基于交互注意力的Transformer融合模块引导双域特征融合,通过挖掘双域特征的相关性,实现跨域信息融合.实验证明,在公开的数据集上,本方法相较于其他基线重建方法均能取得更为优异的重建效果.同时,消融实验证明了本文所提出的网络模块的有效性.

高效跨域的Transformer小样本语义分割网络

小样本语义分割旨在仅使用数个标注样本学习目标类别特征并完成分割任务。主流研究存在的主要问题是:训练效率低下,训练和测试在同一数据域。为此构建了一种基于Transformer的高效、跨域的小样本语义分割网络SGFNet。在编码层,使用共享权重的MixVisionTransformer构建孪生网络,用于提取支持集和查询集的图像特征;在关系计算层,通过计算支持集图像特征向量与其对应mask的哈达玛积,提取目标类别的高维特征,并与查询集图像特征进行关系计算;在解码层,改进基于MLP的解码器,提出了残差解码器,将不同层级的特征解码得到最终分割结果。实验表明,该模型只需要在FSS-1000数据集上使用单张3090 GPU训练1.5~4.0 h,即可在FSS-1000数据集上获得最优结果1-shot mIoU 87.0%,在PASCAL-5i和COCO-20i数据集进行跨域测试达到非跨域的效果,1-shot mIoU分别为60.4%和33.0%,证明了该模型高效且跨域。

小波分频自注意力Transformer图像去雨网络

针对视觉Transformer对高频信息捕捉能力弱以及目前许多图像去雨方法易丢失细节的问题,提出小波分频自注意力Transformer图像去雨网络(WFDST-Net)。小波分频自注意力Transformer(WFDST)作为WFDST-Net的主要模块,其利用不可分提升小波变换获取特征图的低频分量和高频分量,分别在低频和高频中进行自注意力交互,使模块从低频中学习恢复全局结构的能力,在高频中强化捕捉雨纹等线条细节的能力,增强对不同频域特征的建模能力。WFDST-Net采用U形架构并通过不可分提升小波变换获取多尺度特征,可在捕获不同形状高频雨纹的同时保证信息的完整性。相比其他图像去雨相关的Transformer,WFDST-Net具有更低的参数量。此外,提出VOCRain250数据集用于联合图像去雨和语义分割任务,该数据集比目前广泛使用的BDD150更具优势。实验表明,所提方法增强了视觉Transformer对不同频域信息的捕获能力,并在合成和真实数据集以及VOCRain250中的表现优于目前先进的去雨方法,能有效去除复杂雨纹并保留更多细节特征。

基于Transformer块的混合域网络稀疏角度CT成像

目的 针对计算机断层扫描(Computed Tomography,CT)中由于不完全扫描数据导致图像噪声伪影严重的问题,提出一种基于Transformer块的混合域网络稀疏角度CT成像算法(Hybrid Domain network for sparse view CT imaging based on Transformer,HDTransformer)。方法 算法的主要思想是借助于新型的Transformer网络,构建适用于多阶段稀疏角度CT投影数据及图像数据的处理流,以提高稀疏角度CT图像重建质量;与现有两阶段混合域处理方法相比,本方法采用图像域-投影域-图像域三阶段混合处理流程,通过多阶段信息的联合互补提高成像质量;此外,针对不同阶段数据噪声伪影特点设计不同的Transformer块,以实现差异化的处理;更进一步,算法采用可微分的解析重建和投影运算,建立投影域与图像域数据的转换,最终实现端到端的稀疏角度CT优质成像流。结果 通过Mayo数据实验验证,其视觉结果表明:处理后的不同部位CT图像噪声伪影均能够得到较好的抑制;量化结果表明:处理后的CT图像峰值信噪比和特征相似性均优于对比方法。结论 实验的定性和定量结果表明:所提算法在去除图像伪影噪声方面要优于其他算法,具有更高的质量,验证了该方法的有效性。

基于复数域Transformer-Unet混合模型的PolSAR地物分类

传统的基于深度学习的极化合成孔径雷达(Polarimetric Synthetic Aperture Radar,PolSAR)地物分类方法,通过堆叠卷积层提取图像局部特征,难以建立长距离依赖关系。基于自注意力机制的深度学习模型Transformer (变换)在图像分类任务中取得了成功,其自注意力机制能够捕获全局像素之间的关联性,同时PolSAR地物分类任务被证实:相比于实数域,其在复数域上表现出更好的分类效果。因此,本文将Transformer引入到复数域中,提出了一种基于复数域的Transformer和Unet (语义分割网络)混合模型(CT-Unet)用于PolSAR地物分类,将Transformer与CNN相结合,对复数类型的PolSAR数据进行特征提取,使用西安数据集和德国数据集进行PolSAR地物分类的实验结果表明:提出的模型能够有效提高PolSAR地物分类的准确性,Transformer有望在PolSAR地物分类任务中弥补卷积神经网络的不足。

Transformer在域适应中的应用研究综述

作为迁移学习的重要分支,域适应旨在解决传统机器学习算法在训练样本和测试样本服从不同数据分布时性能急剧下降的问题。Transformer是基于自注意力机制的深度学习框架,具有强大的全局特征提取能力和建模能力,近年来Transformer与域适应相结合也成为研究的热点。虽然已有大量相关方法问世,但Transformer应用在域适应的综述却未见报道。为了填补这个领域的空白,为相关研究提供借鉴和参考,对近年来出现的一些基于Transformer的典型域适应方法进行归纳总结与分析,概述域适应的相关概念与Transformer的基本结构,从图像分类、图像语义分割、目标检测、医学图像分析四个应用梳理了各种基于Transformer的域适应方法,对图像分类下的域适应方法进行比较,总结当前域适应Transformer模型存在的挑战并探讨未来可行的研究方向。

APK-CNN和Transformer增强的多域虚假新闻检测模型

为解决社交媒体新闻中的领域转移、领域标签不完整问题,以及探索更高效的多域新闻文本特征提取和融合网络,提出一种基于APK-CNN(Adaptive Pooling Kernel Convolutional Neural Network)和Transformer增强的多域虚假新闻检测模型Transm3。首先,设计三通道网络对文本的语义、情感和风格信息进行特征提取和表示,并利用多粒度跨域交互器对这些特征进行视图组合;其次,通过优化的软共享内存网络和域适配器来完善新闻领域标签;再次,将Transformer与多粒度跨域交互器结合,使用更先进的融合网络动态加权聚合不同领域的交互特征;最后,将融合特征输入分类器中用于真/假新闻判别。实验结果表明,Transm3与M3FEND(Memory-guided Multi-view Multi-domain FakE News Detection)和EANN(Event Adversarial Neural Networks for multi-modal fake news detection)相比,综合F1值在中文数据集上分别提高了3.68%和6.46%,在英文数据集上分别提高了6.75%和11.93%,在各分领域上F1值也有明显的提高,充分验证了Transm3在多域虚假新闻检测工作上的有效性。

Performance of Wavelet-Transform-Domain Adaptive Equalizers

In this paper performances of wavelet transform domain (WTD) adaptive equalizers based on the least mean￣square (LMS) algorithm are analyzed. The optimum Wiener solution, the condition of convergence, the minimum mean square error (MSE) and the steady state excess MSE of the WTD adaptive equalizer are obtained. Constant and time varying convergence factor adaptive algorithms are studied respectively. Computational complexities of WTD LMS equalizers are given. The equalizer in WTD shows much better convergence performance than that of the conventional in time domain.

Efficient reconstruction scheme for transform domain Wyner-Ziv video coding

To minimize the errors of the reconstructed values and improve the quality of decoded image, an efficient reconstruction scheme for transform domain Wyner-Ziv （WZ） video coding is proposed. The reconstruction scheme exploits temporal correlation of the coefficient bands, the WZ decoded bits stream and the side information efficiently. When side information is outside the decoded quantization bin, the reconstructed value is derived using expectation of the WZ decoded bit stream and the side information. When side information is within the decoded quantization bin, the reconstructed value is derived using the biased predictor. Simulation results show that the proposed reconstruction scheme gains up to 1.32 dB compared with the commonly used boundary reconstruction scheme at the same bit rates and similar computation cost.

Post-Mesozoic Transformation of Tectonic Domain in Southeastern China and Its Geodynamic Mechanism

Since the Mesozoic and Cenozoic, a transformation from a Tethyan Himalayan tectonic domain into a circum Pacific tectonic domain from Indosinian to Yanshanian is indicated in this paper, resulting in conspicuous changes in geophysics, tectono magmatic distribution, lithofacies and paleo geography, tectonic system in southeastern China. Tectonic analysis shows that the tectonic framework resulted from the compounding, transforming and superimposing of the two tectonic domains. The geodynamic mechanism of the transformation is mainly shown as the transverse and longitudinal heterogeneity of lithosphere, and the exchange between the crust and the mantle.

基于多尺度Transformer融合多域信息的伪造人脸检测

当前,基于Deepfakes等深度伪造技术生成的“换脸”类伪造视频泛滥,给公民个人隐私和国家政治安全带来巨大威胁,为此,研究视频中深度伪造人脸检测技术具有重要意义。针对已有伪造人脸检测方法存在的面部特征提取不充分、泛化能力弱等不足,提出一种基于多尺度Transformer对多域信息进行融合的伪造人脸检测方法。基于多域特征融合的思路,同时从视频帧的频域与RGB域进行特征提取,提高模型的泛化性;联合EfficientNet和多尺度Transformer,设计多层级的特征提取网络以提取更精细的伪造特征。在开源数据集上的测试结果表明,相比已有方法,所提方法具有更好的检测效果;同时在跨数据集上的实验结果证明了所提模型具有较好的泛化性能。

基于改进Transformer和强化学习的僵尸网络DGA域名检测

针对现有僵尸网络检测方法检测精度不高和检测时间开销较大的问题,提出一种基于改进Transformer和强化学习的僵尸网络域名生成算法(Domain Generation Algorithm,DGA)的域名检测方法。首先,利用深度可分离卷积替换ResNet和ResNeXt网络中的卷积块,通过减少网络模型参数来降低模型的时间开销;其次,利用改进后的ResNet和ResNeXt网络将域名字符串映射到深度特征空间,构造多尺度特征,强化特征的表达能力;再次,利用长短期记忆神经网络(Long Short-Term Memory,LSTM)对Transformer网络进行改进,在保持字符间相对位置的同时,进一步建立上下文的长距离依赖编码,并在此基础上引入注意力机制,强化模型对关键特征的捕获能力;最后,引入强化学习对模型进行微调,提高DGA域名的检测精度。在多个DGA域名数据集上进行测试验证,结果表明该模型在保持检测时间开销较小的基础上,具有更高的检测精度。

Power System Harmonic Detection Using Frequency-Domain Interpolation Wavelet Transform

Aiming at harmonic detection, fast Fourier transform can only detect integer harmonics precisely, short time Fourier transform can detect non-integer harmonics with low resolution, and some former wavelet based methods have no aliasing-reduction scheme which result in low measurement precision and poor robustness. A frequency-domain interpolation algorithm to detect harmonics is proposed by choosing Shannon wavelet. Shannon wavelet is an orthogonal wavelet possessing best ideal frequency domain localization ability, it can restrict wavelet abasing but bring about Gibbs oscillation phenomenon simultaneously. An interpolation algorithm is developed to overcome this problem. Simulation reveals that the proposed method can effectively cancel aliasing, spectral leakage and Gibbs phenomenon, so it provides an effective means for power system harmonic analysis.

基于频域Transformer的对抗生成网络去运动模糊算法

对于图像不均匀的运动模糊问题,模型感受野的大小相当重要,基于自注意力机制的Transformer的感受野远大于传统CNN模型。并且引入了高效的频域处理模块,以往结合频域处理模糊的方法忽略了运动模糊在频域图中呈现出来的方向性特征,根据这种特征将其分解处理可以高效地恢复图像。实验中与目前较经典的几个方法进行了对比,模型在GoPro数据集上PSNR得分为32.24 dB,在RealBlur-J数据集上的得分为29.38 dB,效果相较于其他方法在两个数据集上分别提高了0.72 dB和0.83 dB。