Non-linear Convolution Based Controllable Image Analogies and Self Analogies

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.