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
We present new connections among linear anomalous diffusion (AD), normal diffusion (ND) and the Central Limit Theorem (CLT). This is done by defining a point transformation to a new position variable, which we postula...We present new connections among linear anomalous diffusion (AD), normal diffusion (ND) and the Central Limit Theorem (CLT). This is done by defining a point transformation to a new position variable, which we postulate to be Cartesian, motivated by considerations from super-symmetric quantum mechanics. Canonically quantizing in the new position and momentum variables according to Dirac gives rise to generalized negative semi-definite and self-adjoint Laplacian operators. These lead to new generalized Fourier transformations and associated probability distributions, which are form invariant under the corresponding transform. The new Laplacians also lead us to generalized diffusion equations, which imply a connection to the CLT. We show that the derived diffusion equations capture all of the Fractal and Non-Fractal Anomalous Diffusion equations of O’Shaughnessy and Procaccia. However, we also obtain new equations that cannot (so far as we can tell) be expressed as examples of the O’Shaughnessy and Procaccia equations. The results show, in part, that experimentally measuring the diffusion scaling law can determine the point transformation (for monomial point transformations). We also show that AD in the original, physical position is actually ND when viewed in terms of displacements in an appropriately transformed position variable. We illustrate the ideas both analytically and with a detailed computational example for a non-trivial choice of point transformation. Finally, we summarize our results.展开更多
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We present new connections among linear anomalous diffusion (AD), normal diffusion (ND) and the Central Limit Theorem (CLT). This is done by defining a point transformation to a new position variable, which we postulate to be Cartesian, motivated by considerations from super-symmetric quantum mechanics. Canonically quantizing in the new position and momentum variables according to Dirac gives rise to generalized negative semi-definite and self-adjoint Laplacian operators. These lead to new generalized Fourier transformations and associated probability distributions, which are form invariant under the corresponding transform. The new Laplacians also lead us to generalized diffusion equations, which imply a connection to the CLT. We show that the derived diffusion equations capture all of the Fractal and Non-Fractal Anomalous Diffusion equations of O’Shaughnessy and Procaccia. However, we also obtain new equations that cannot (so far as we can tell) be expressed as examples of the O’Shaughnessy and Procaccia equations. The results show, in part, that experimentally measuring the diffusion scaling law can determine the point transformation (for monomial point transformations). We also show that AD in the original, physical position is actually ND when viewed in terms of displacements in an appropriately transformed position variable. We illustrate the ideas both analytically and with a detailed computational example for a non-trivial choice of point transformation. Finally, we summarize our results.
