目的近年来,基于深度学习的水印方法得到了广泛研究。现有方法通常对特征图的低频和高频部分同等对待,忽视了不同频率成分之间的重要差异,导致模型在处理多样化攻击时缺乏灵活性,难以同时实现水印的高保真性和强鲁棒性。为此,本文提出...目的近年来,基于深度学习的水印方法得到了广泛研究。现有方法通常对特征图的低频和高频部分同等对待,忽视了不同频率成分之间的重要差异,导致模型在处理多样化攻击时缺乏灵活性,难以同时实现水印的高保真性和强鲁棒性。为此,本文提出一种频率感知驱动的深度鲁棒图像水印技术(deep robust image watermarking driven by frequency awareness,RIWFP)。方法通过差异化机制处理低频和高频成分,提升水印性能。具体而言,低频成分通过小波卷积神经网络进行建模,利用宽感受野卷积在粗粒度层面高效学习全局结构和上下文信息;高频成分则采用深度可分离卷积和注意力机制组成的特征蒸馏块进行精炼,强化图像细节,在细粒度层面高效捕捉高频信息。此外,本文使用多频率小波损失函数,引导模型聚焦于不同频带的特征分布,进一步提升生成图像的质量。结果实验结果表明,提出的频率感知驱动的深度鲁棒图像水印技术在多个数据集上均表现出优越性能。在COCO(common objects in context)数据集上,RIWFP在随机丢弃攻击下的准确率达到91.4%;在椒盐噪声和中值滤波攻击下,RIWFP分别以100%和99.5%的准确率达到了最高水平,展现了其对高频信息的高效学习能力。在Ima⁃geNet数据集上,RIWFP在裁剪攻击下的准确率为93.4%;在JPEG压缩攻击下的准确率为99.6%,均显著优于其他对比方法。综合来看,RIWFP在COCO和ImageNet数据集上的平均准确率分别为96.7%和96.9%,均高于其他对比方法。结论本文所提方法通过频率感知的粗到细处理策略,显著增强了水印的不可见性和鲁棒性,在处理多种攻击时表现出优越性能。展开更多
Marriage has become less common,while the incidence of divorce has risen in Iran.These have made marriage facilitation and divorce prevention as the cornerstone of population policy.It is clear that prediction of the ...Marriage has become less common,while the incidence of divorce has risen in Iran.These have made marriage facilitation and divorce prevention as the cornerstone of population policy.It is clear that prediction of the incidence of marriage and divorce will help policy makers to design effective interventions.This paper uses the number of marriages and divorces between 1980 and 2017,published by the Statistical Center of Iran,to predict the incidence of marriage and divorce through 2027 at the national level.Given the limitations of common methods,such as ARMAX,ARMA,MR and AR,in predicting time series with abrupt changes,this paper applies a mixed method,which combines the Neural Network and the Wavelet mathematical tools.The comparison between the data and the results obtained from the wavelet-neural network confirms the precision of the model.The precision and validity of the neural-wavelet network model is further confirmed by the fact that it has been able to reduce the mean sum of square of errors to a larger extent than the Neural Network models.The findings show a 3%decrease in the number of marriages,from 704,716 in 2017 to 683,190 in 2027.On the other hand,the number of divorces has increased by 30%,from 181,049 in 2017 to 235,407 in 2027.Thus,the number of divorces per 100 marriages will increase from 25.7 to 34.5 just in a decade,which calls for effective interventions if family formation and consolidation are to be improved in Iran.展开更多
文摘目的近年来,基于深度学习的水印方法得到了广泛研究。现有方法通常对特征图的低频和高频部分同等对待,忽视了不同频率成分之间的重要差异,导致模型在处理多样化攻击时缺乏灵活性,难以同时实现水印的高保真性和强鲁棒性。为此,本文提出一种频率感知驱动的深度鲁棒图像水印技术(deep robust image watermarking driven by frequency awareness,RIWFP)。方法通过差异化机制处理低频和高频成分,提升水印性能。具体而言,低频成分通过小波卷积神经网络进行建模,利用宽感受野卷积在粗粒度层面高效学习全局结构和上下文信息;高频成分则采用深度可分离卷积和注意力机制组成的特征蒸馏块进行精炼,强化图像细节,在细粒度层面高效捕捉高频信息。此外,本文使用多频率小波损失函数,引导模型聚焦于不同频带的特征分布,进一步提升生成图像的质量。结果实验结果表明,提出的频率感知驱动的深度鲁棒图像水印技术在多个数据集上均表现出优越性能。在COCO(common objects in context)数据集上,RIWFP在随机丢弃攻击下的准确率达到91.4%;在椒盐噪声和中值滤波攻击下,RIWFP分别以100%和99.5%的准确率达到了最高水平,展现了其对高频信息的高效学习能力。在Ima⁃geNet数据集上,RIWFP在裁剪攻击下的准确率为93.4%;在JPEG压缩攻击下的准确率为99.6%,均显著优于其他对比方法。综合来看,RIWFP在COCO和ImageNet数据集上的平均准确率分别为96.7%和96.9%,均高于其他对比方法。结论本文所提方法通过频率感知的粗到细处理策略,显著增强了水印的不可见性和鲁棒性,在处理多种攻击时表现出优越性能。
文摘Marriage has become less common,while the incidence of divorce has risen in Iran.These have made marriage facilitation and divorce prevention as the cornerstone of population policy.It is clear that prediction of the incidence of marriage and divorce will help policy makers to design effective interventions.This paper uses the number of marriages and divorces between 1980 and 2017,published by the Statistical Center of Iran,to predict the incidence of marriage and divorce through 2027 at the national level.Given the limitations of common methods,such as ARMAX,ARMA,MR and AR,in predicting time series with abrupt changes,this paper applies a mixed method,which combines the Neural Network and the Wavelet mathematical tools.The comparison between the data and the results obtained from the wavelet-neural network confirms the precision of the model.The precision and validity of the neural-wavelet network model is further confirmed by the fact that it has been able to reduce the mean sum of square of errors to a larger extent than the Neural Network models.The findings show a 3%decrease in the number of marriages,from 704,716 in 2017 to 683,190 in 2027.On the other hand,the number of divorces has increased by 30%,from 181,049 in 2017 to 235,407 in 2027.Thus,the number of divorces per 100 marriages will increase from 25.7 to 34.5 just in a decade,which calls for effective interventions if family formation and consolidation are to be improved in Iran.
文摘针对传统滚动轴承故障诊断方法过度依赖人工提取与分析特征、模型泛化性差以及对时序和通道深层次特征读取不充分的问题,提出了一种基于时频图与改进的卷积神经网络(Convolutional Neural Network,CNN)相结合的滚动轴承故障诊断方法。首先,将滚动轴承的原始振动信号经过连续小波变换(Continuous Wavelet Transform,CWT)转化为二维时频图,再利用内嵌长短期记忆网络(Long Short Term Memory,LSTM)的二维卷积神经网络从变换后的时频图中充分提取图像的时序特征,然后,通过高效通道注意力机制(Efficient Channel Attention,ECA)获取通道的全局信息并自适应地对各通道权重值进行动态调整,建立通道间的联系,自适应提取深层次关键特征。最后,利用凯斯西储大学滚动轴承故障数据集进行实验验证。实验结果表明,相较于一些常见的滚动轴承故障诊断方法,该方法在诊断准确率方面有明显提高。
