From the view of information flow, a super-network equilibrium optimization model is proposed to compute the solution of the operation architecture which is made up of a perceptive level, a command level and a firepow...From the view of information flow, a super-network equilibrium optimization model is proposed to compute the solution of the operation architecture which is made up of a perceptive level, a command level and a firepower level. Firstly, the optimized conditions of the perceptive level, command level and firepower level are analyzed respectively based on the demand of information relation,and then the information supply-and-demand equilibrium model of the operation architecture super-network is established. Secondly,a variational inequality transformation(VIT) model for equilibrium optimization of the operation architecture is given. Thirdly, the contraction projection algorithm for solving the operation architecture super-network equilibrium optimization model with fuzzy demands is designed. Finally, numerical examples are given to prove the validity and rationality of the proposed method, and the influence of fuzzy demands on the super-network equilibrium solution of operation architecture is discussed.展开更多
针对现有生成对抗网络的单图像超分辨率重建在大尺度因子下存在训练不稳定、特征提取不足和重建结果纹理细节严重缺失的问题,提出一种拆分注意力网络的单图超分辨率重建方法。首先,以拆分注意力残差模块作为基本残差块构造生成器,提高...针对现有生成对抗网络的单图像超分辨率重建在大尺度因子下存在训练不稳定、特征提取不足和重建结果纹理细节严重缺失的问题,提出一种拆分注意力网络的单图超分辨率重建方法。首先,以拆分注意力残差模块作为基本残差块构造生成器,提高生成器特征提取的能力。其次,在损失函数中引入鲁棒性更好的Charbonnier损失函数和Focal Frequency Loss损失函数代替均方差损失函数,同时加入正则化损失平滑训练结果,防止图像过于像素化。最后,在生成器和判别器中采用谱归一化处理,提高网络的稳定性。在4倍放大因子下,与其他方法在Set5、Set14、BSDS100、Urban100测试集上进行测试比较,本文方法的峰值信噪比比其他对比方法的平均值提升1.419 dB,结构相似性比其他对比方法的平均值提升0.051。实验数据和效果图表明,该方法主观上具有丰富的细节和更好的视觉效果,客观上具有较高的峰值信噪比值和结构相似度值。展开更多
传统的基于对比学习的图像超分辨率方法,一般将原始图像作为正样本,将退化图像或其他类图像作为负样本,存在对纹理细节恢复差的问题。本文提出基于对比学习的深度残差网络图像超分辨率(depth residual image super-resolution based on ...传统的基于对比学习的图像超分辨率方法,一般将原始图像作为正样本,将退化图像或其他类图像作为负样本,存在对纹理细节恢复差的问题。本文提出基于对比学习的深度残差网络图像超分辨率(depth residual image super-resolution based on contrast learning,CEDSR)方法,针对残差超分辨率模型,采用对高分辨率图像锐化后的图像作为正样本,对高分辨率图像轻微模糊的图像作为负样本,利用正负样本下的对比损失提升对纹理细节的恢复增强。增强锐化后的正样本图像携带更丰富的纹理信息,基于不同函数生成的模糊负样本图像刻画了纹理模糊特征,正负样本构建的对比损失有利于图像超分辨率图像对纹理细节的恢复。本文模型在4个标准数据集DIV2K、Set14、BSDS100和Urban100上与经典算法进行实验对比,定性和定量实验结果均表明本文模型可以获得效果更好的超分辨率图像。展开更多
基金supported by the National Natural Science Foundation of China (71771216,71701209)Shaanxi Natural Science Foundation (2019 JQ-250)。
文摘From the view of information flow, a super-network equilibrium optimization model is proposed to compute the solution of the operation architecture which is made up of a perceptive level, a command level and a firepower level. Firstly, the optimized conditions of the perceptive level, command level and firepower level are analyzed respectively based on the demand of information relation,and then the information supply-and-demand equilibrium model of the operation architecture super-network is established. Secondly,a variational inequality transformation(VIT) model for equilibrium optimization of the operation architecture is given. Thirdly, the contraction projection algorithm for solving the operation architecture super-network equilibrium optimization model with fuzzy demands is designed. Finally, numerical examples are given to prove the validity and rationality of the proposed method, and the influence of fuzzy demands on the super-network equilibrium solution of operation architecture is discussed.
文摘针对现有生成对抗网络的单图像超分辨率重建在大尺度因子下存在训练不稳定、特征提取不足和重建结果纹理细节严重缺失的问题,提出一种拆分注意力网络的单图超分辨率重建方法。首先,以拆分注意力残差模块作为基本残差块构造生成器,提高生成器特征提取的能力。其次,在损失函数中引入鲁棒性更好的Charbonnier损失函数和Focal Frequency Loss损失函数代替均方差损失函数,同时加入正则化损失平滑训练结果,防止图像过于像素化。最后,在生成器和判别器中采用谱归一化处理,提高网络的稳定性。在4倍放大因子下,与其他方法在Set5、Set14、BSDS100、Urban100测试集上进行测试比较,本文方法的峰值信噪比比其他对比方法的平均值提升1.419 dB,结构相似性比其他对比方法的平均值提升0.051。实验数据和效果图表明,该方法主观上具有丰富的细节和更好的视觉效果,客观上具有较高的峰值信噪比值和结构相似度值。
文摘传统的基于对比学习的图像超分辨率方法,一般将原始图像作为正样本,将退化图像或其他类图像作为负样本,存在对纹理细节恢复差的问题。本文提出基于对比学习的深度残差网络图像超分辨率(depth residual image super-resolution based on contrast learning,CEDSR)方法,针对残差超分辨率模型,采用对高分辨率图像锐化后的图像作为正样本,对高分辨率图像轻微模糊的图像作为负样本,利用正负样本下的对比损失提升对纹理细节的恢复增强。增强锐化后的正样本图像携带更丰富的纹理信息,基于不同函数生成的模糊负样本图像刻画了纹理模糊特征,正负样本构建的对比损失有利于图像超分辨率图像对纹理细节的恢复。本文模型在4个标准数据集DIV2K、Set14、BSDS100和Urban100上与经典算法进行实验对比,定性和定量实验结果均表明本文模型可以获得效果更好的超分辨率图像。
