Research On Fast Style Transfer Algorithms Based On Deep Learning

Style Transfer aims at repainting the content image with the texture of the style image while retaining the semantic information of the content image.It can not only meet people's needs for beauty and art,but also can be applied in generation of art painting video production,game scene design and so on.Style Transfer accelerates the creation of art works with artificial intelligence and can be widely used in many applications.Style Transfer based on deep learning has made great progress.However,there still exist some shortages.Most of them can not generate high-resolution stylized images with high quality and suffer from huge model parameters and low computational efficiency.To solve these problems,we conduct the study of Style Transfer based on deep learning.The main contents are as follows:(1)A faster style transfer network with sub-pixel convolution is proposed.Most of the Per-Style-Per-Model Fast Neural Method can not balance the quality and speed of conversion on high resolution images.Our algorithm takes a low-resolution content image as input,and generates the high-resolution stylized image by sub-pixel convolution and dense connection with Post-normalization.Dense connection captures multi-scale features and reuse these features well,so it requires less parameters.Experiments demonstrate that our algorithm achieves better performance with faster speed on high-resolution images.(2)An Arbitrary-Style-Per-Model Fast Neural Method based on multi-scale fusion and compress attention is proposed.Most models of Arbitrary-Style-Per-Model Fast Neural Method can not handle multi-scale features efficiently and suffer from huge model parameters.The multi-scale fusion and compress attention we proposed can handle multi-scale features well and only use one attention module to capture multi-scale features from the encoder module.Besides we compress the features with multi-scale to make it more efficiently.We also design a lighter encoder to encode the content and style images.Experiments demonstrate that our algorithm can generate stylized images with higher quality and less computation under the same conditions.The research contents take into account the quality and efficiency of Style Transfer,which have great application value.