Weak pairing image style migration method based on pose self-supervised generative adversarial network

Abstract

The invention discloses a weak pairing image style migration method based on a pose self-supervised generative adversarial network, and belongs to the field of image processing. The method is suitable for style migration of weak pairing images, different style pictures with certain overlap are utilized to carry out model training of the adversarial neural network, so that the model is insensitive to poses and focuses on style learning, and therefore, in the actual application process, a source style can be converted into a target style, but the poses are kept unchanged. In addition, in the model training process of the adversarial neural network, a differentiable pose solver capable of estimating the relative pose of any two images is introduced, a phase correlation algorithm is optimized to be differentiable, and the phase correlation algorithm is embedded into an end-to-end learning network framework to achieve pose estimation. According to the method, style migration of weak pairing data can be realized, and support is provided for a robot self-positioning technology.

Description

technical field [0001] The invention belongs to the field of computer vision style transfer, and in particular relates to a style transfer of pictures through a pose self-supervised confrontation generation network. Background technique [0002] Image style transfer endows the robot with a higher-level understanding of the environment and helps the robot adapt to different scenarios. Therefore, tasks trained or configured on one scene can be easily performed on other scenes through this transfer learning. This setup is extremely helpful for a variety of robotic tasks, such as transferring detectors trained on sunny days to rainy nights, place re-identification, and semantic segmentation across domains. [0003] Currently, most works focus on fully paired image style transfer or fully unpaired image style transfer during the training process of image style transfer networks. In paired image transfer, the content of two images from two styles is exactly the same. To handle ...

AI Technical Summary

Problems solved by technology

[0010] The purpose of the present invention is to solve the problem that the style cannot be transferred during the style transfer in the prior art without changing the pose of the original image, and to provide a weakly paired image style transfer method based on the pose self-supervised confrontation generation network

Images