结合可逆网络的超高分辨率图像的风格迁移

针对处理超高分辨率图像时面临的内存成本和风格迁移过程中过度风格化，提出了一种结合可逆网络的超高分辨率图像的风格迁移方法。该算法采用可逆的Glow模块作为基本单元构建可逆神经网络模块，并将图像分为小块处理；在风格迁移模块中使用具有通道注意力机制的残差模块和缩略图实例化归一化模块(TIN),以保证各模块风格一致；提出基于全局-局部的损失计算方式，能够有效地处理局部的结构特征。实验结果表明，相较于当前通用的各种神经风格迁移网络，所提算法不仅能够避免图像在编码和解码过程中的信息丢失问题，而且能以更低的内存成本实现更优的风格迁移。

Super-Resolution Image Style Transfer Combined with Reversible Network

Aiming at the problem of memory consumption in ultra-high resolution image processing and the problem of over-stylization in the process of style transfer, a method of ultra-high resolution image style transfer combined with reversible network is proposed. The algorithm used the reversible Glow module as the basic unit to construct a reversible neural network module, and divided the image into small blocks for processing. In the style transfer module, a residual module with a channel attention mechanism and a thumbnail instantiation normalization module (TIN) were used to ensure that the styles of each module were consistent. A global-local loss calculation method was proposed, which could effectively deal with local structural features. Experimental results show that, compared with the current general-purpose neural style transfer network, this algorithm can not only avoid the information loss problem in the process of image encoding and decoding, but also achieve better style transfer performance at a lower memory cost.