针对现有深度学习算法在壁画修复时,存在全局语义一致性约束不足及局部特征提取不充分,导致修复后的壁画易出现边界效应和细节模糊等问题,提出一种双向自回归Transformer与快速傅里叶卷积增强的壁画修复方法.首先,设计基于Transformer...针对现有深度学习算法在壁画修复时,存在全局语义一致性约束不足及局部特征提取不充分,导致修复后的壁画易出现边界效应和细节模糊等问题,提出一种双向自回归Transformer与快速傅里叶卷积增强的壁画修复方法.首先,设计基于Transformer结构的全局语义特征修复模块,利用双向自回归机制与掩码语言模型(masked language modeling,MLM),提出改进的多头注意力全局语义壁画修复模块,提高对全局语义特征的修复能力.然后,构建了由门控卷积和残差模块组成的全局语义增强模块,增强全局语义特征一致性约束.最后,设计局部细节修复模块,采用大核注意力机制(large kernel attention,LKA)与快速傅里叶卷积提高细节特征的捕获能力,同时减少局部细节信息的丢失,提升修复壁画局部和整体特征的一致性.通过对敦煌壁画数字化修复实验,结果表明,所提算法修复性能更优,客观评价指标均优于比较算法.展开更多
In low earth orbit (LEO) satellite or missile communication scenarios, signals may experience extremely large Doppler shifts and have short visual time. Thus, direct sequence spread spectrum (DSSS) systems should be a...In low earth orbit (LEO) satellite or missile communication scenarios, signals may experience extremely large Doppler shifts and have short visual time. Thus, direct sequence spread spectrum (DSSS) systems should be able to achieve acquisition in a very short time in spite of large Doppler frequencies. However, the traditional methods cannot solve it well. This work describes a new method that uses a differential decoding technique for Doppler mitigation and a batch process of FFT (fast Fourier transform) and IFFT (invert FFT) for the purpose of parallel code phase search by frequency domain correlation. After the code phase is estimated, another FFT process is carried out to search the Doppler frequency. Since both code phase and Doppler frequency domains are searched in parallel, this architecture can provide acquisition fifty times faster than conventional FFT methods. The performance in terms of the probability of detection and false alarm are also analyzed and simulated, showing that a signal-to-noise ratio (SNR) loss of 3 dB is introduced by the differential decoding. The proposed method is an efficient way to shorten the acquisition time with slightly hardware increasing.展开更多
为解决传统特高压直流保护对高阻故障检测准确率不高、故障检测时间过长以及故障选极不完善的问题,提出基于长短时记忆(long short term memory,LSTM)循环神经网络(recurrent neural network,RNN)的特高压直流输电线路继电保护故障检测...为解决传统特高压直流保护对高阻故障检测准确率不高、故障检测时间过长以及故障选极不完善的问题,提出基于长短时记忆(long short term memory,LSTM)循环神经网络(recurrent neural network,RNN)的特高压直流输电线路继电保护故障检测方法。首先,基于快速傅里叶变换分析特高压直流输电系统暂态故障特征,使用相模变换和小波变换提取出故障特征量作为输入数据。其次,将输入数据输入到LSTM-RNN中进行前向传播,对系统故障特征进行深度学习,同时使用反向传播方式更新网络参数,将深层的特征量输入到Softmax分类器中进行分类,把故障识别分成区外故障、母线故障和线路故障,故障分类为正极故障、负极故障和双极故障,并输出识别结果。最后,在PSCAD/EMTDC仿真条件下,搭建特高压直流输电模型。验证结果表明:所提的方法在特高压直流输电线路继电保护的故障检测、故障选极上具有更好的效果,相比于人工神经网络、卷积神经网络、支持向量机,故障识别准确率分别提升4.71%、6.57%、9.32%。展开更多
文摘针对现有深度学习算法在壁画修复时,存在全局语义一致性约束不足及局部特征提取不充分,导致修复后的壁画易出现边界效应和细节模糊等问题,提出一种双向自回归Transformer与快速傅里叶卷积增强的壁画修复方法.首先,设计基于Transformer结构的全局语义特征修复模块,利用双向自回归机制与掩码语言模型(masked language modeling,MLM),提出改进的多头注意力全局语义壁画修复模块,提高对全局语义特征的修复能力.然后,构建了由门控卷积和残差模块组成的全局语义增强模块,增强全局语义特征一致性约束.最后,设计局部细节修复模块,采用大核注意力机制(large kernel attention,LKA)与快速傅里叶卷积提高细节特征的捕获能力,同时减少局部细节信息的丢失,提升修复壁画局部和整体特征的一致性.通过对敦煌壁画数字化修复实验,结果表明,所提算法修复性能更优,客观评价指标均优于比较算法.
基金Project(60904090) supported by the National Natural Science Foundation of China
文摘In low earth orbit (LEO) satellite or missile communication scenarios, signals may experience extremely large Doppler shifts and have short visual time. Thus, direct sequence spread spectrum (DSSS) systems should be able to achieve acquisition in a very short time in spite of large Doppler frequencies. However, the traditional methods cannot solve it well. This work describes a new method that uses a differential decoding technique for Doppler mitigation and a batch process of FFT (fast Fourier transform) and IFFT (invert FFT) for the purpose of parallel code phase search by frequency domain correlation. After the code phase is estimated, another FFT process is carried out to search the Doppler frequency. Since both code phase and Doppler frequency domains are searched in parallel, this architecture can provide acquisition fifty times faster than conventional FFT methods. The performance in terms of the probability of detection and false alarm are also analyzed and simulated, showing that a signal-to-noise ratio (SNR) loss of 3 dB is introduced by the differential decoding. The proposed method is an efficient way to shorten the acquisition time with slightly hardware increasing.
文摘为解决传统特高压直流保护对高阻故障检测准确率不高、故障检测时间过长以及故障选极不完善的问题,提出基于长短时记忆(long short term memory,LSTM)循环神经网络(recurrent neural network,RNN)的特高压直流输电线路继电保护故障检测方法。首先,基于快速傅里叶变换分析特高压直流输电系统暂态故障特征,使用相模变换和小波变换提取出故障特征量作为输入数据。其次,将输入数据输入到LSTM-RNN中进行前向传播,对系统故障特征进行深度学习,同时使用反向传播方式更新网络参数,将深层的特征量输入到Softmax分类器中进行分类,把故障识别分成区外故障、母线故障和线路故障,故障分类为正极故障、负极故障和双极故障,并输出识别结果。最后,在PSCAD/EMTDC仿真条件下,搭建特高压直流输电模型。验证结果表明:所提的方法在特高压直流输电线路继电保护的故障检测、故障选极上具有更好的效果,相比于人工神经网络、卷积神经网络、支持向量机,故障识别准确率分别提升4.71%、6.57%、9.32%。
