Image expansion method based on perception loss and style loss

Abstract

The invention relates to an image expansion method based on perception loss and style loss, and the method comprises the steps: transmitting a preprocessed data set into a constructed image expansionnetwork which comprises a reconstruction path and a generation path, wherein the reconstruction path is used for inputting a to-be-completed region image, obtaining the prior information of a to-be-completed part, and finally reconstructing an original image, and the generation path is used for inputting a missing image. The priori distribution obtained by the reconstruction path is used for guiding the generation process of the image, perception loss and style loss constraints are mainly introduced to generate the texture and style of the image in the training process, and a distorted fuzzy structure of a traditional method is improved. Semantic information and an overall style of a known area are obtained through perception loss and style loss respectively, and therefore the network cangrasp the real texture style of the image.

Description

technical field [0001] The invention relates to the field of image processing, in particular to an image extension method for perceptual loss and style loss. Background technique [0002] In computer vision tasks, image completion based on deep learning is widely used, especially in recent years, image completion tasks have achieved remarkable results. Image completion is actually a special task between image editing and image generation. At present, there are several typical methods for processing image completion: one is a simple texture repair method, which only collects similar pixels from existing pixels to fill in the missing area. Since the sample is directly sampled from the image, the connection of the filled part is unnatural. The effect is very poor; the second is to adopt the idea of ​​data-driven, using the data set to learn the distribution of relevant data, which can produce a logical structure, but the repair area is often blurred; there is also a generative...

AI Technical Summary

Problems solved by technology

Its shortcomings are: its memory requirement is relatively large and it only learns the internal block information of the image, which is too dependent on the surrounding information of the image. When the content of the image is missing too much, it cannot make full use of the adjacent information, so the final output is blurred. unnatural situation

[0007] More importantly, in the existing technology, there are more studies on image inward expansion, but less research on image external expansion, and image external expansion is much more difficult than image internal expansion, because the prior information that internal expansion can use more, and the uncertainty of expansion is greater

Images