Rapid image style migration method based on group normalization

Abstract

The invention discloses a rapid image style migration method based on group normalization, and the method comprises the steps: inputting an input image x into a generator network, generating an imageyk, and outputting the image yk; inputting the output image yk and the original content image yc into a loss network, and calculating to obtain a content loss function; inputting the output image yk and the original style image ys into a loss network, and calculating to obtain a style loss function; linearly combining the content loss function and the style loss function into a new loss function as a total loss function; using an MS-COCO data set as a content image for training; and obtaining a result image after style migration. According to the method, batch normalization is replaced by group normalization, the performance of the deep neural network is remarkably improved, the convergence speed of the model is increased in image style migration work, a high-quality image is obtained, anda framework combining a generator network and a loss network is adopted when the rapid image style migration model is constructed, so that the purpose of real-time performance is achieved.

Description

Technical field [0001] The invention belongs to the technical field of image processing, and in particular relates to a fast image style transfer method based on group normalization. Background technique [0002] Image style transfer refers to the use of algorithms to learn the style features of a painting, and then apply this style feature to another picture. Before 2015, most image style transfer methods were realized by manually establishing physical models, mainly using mathematical modeling and mathematical formula induction to generate texture. Efros et al. proposed a simple texture synthesis algorithm that stitched together small pieces of existing texture images and then synthesized new texture images. Hertzmann et al. proposed a new framework for image analogy, which achieved good results in texture synthesis and texture migration. Qian Xiaoyan and others improved the classic Ashikhmin algorithm and proposed a new neighborhood measurement method, which effectively real...

AI Technical Summary

Problems solved by technology

With the rapid development of mobile phone digital functions, the pixels of mobile phone cameras are getting higher and higher, and the resolution of the photos taken is also improved. In the case of inputting high-resolution photos, the resulting pictures obtained by the previous methods often appear Blurring, the resolution of the photo taken by the mobile phone is inconsistent, and the color layout of the converted picture will be inconsistent with the original style picture, and the extraction of fine texture features is not obvious

Because it is applied to mobile phone filters, it is necessary to provide more different types of styles for people to choose from. The previous method converges slowly in training models of different styles, and the time and cost are too large, which restricts the actual application.

Images