基于非线性卷积的可控图像类推和自类推技术

Non-linear Convolution Based Controllable Image Analogies and Self Analogies

提出一种新的图像类推技术,定义"风格"为作用于内容上的非线性卷积过程.通过估计卷积核实现风格学习,通过执行卷积实现风格的传递.该定义能够极大地加快类推速度,提高风格学习与样本数据的独立性,增强图像类推的稳定性与适用性.同时还提出了对图像进行迭代式风格化作用的连续类推思想,以及通过自身构建训练集的自类推思想.连续类推可以生成不同强度的风格化序列,进而实现图像类推控制;图像自类推则可以应用于超分辨等与尺度相关的问题.

This paper presents a novel method of image analogies, to promote the learning from image style. By a definition of style as a non-linear convolution, the style of the image may be studied by kernel estimation, and style transfer could be conducted by executing the learned convolution. Furthermore, we may repeatedly apply a stylization process on an image to generate an analogized image series, in help to realize controllable image analogies. In addition, we can also extend this idea to self analogies, applied to enhance the sharpness of images.