Semantic reconstruction face super-division method and device based on deep reinforcement learning

Abstract

The invention discloses a semantic reconstruction face super-division method and device based on deep reinforcement learning, and the method comprises the following steps: carrying out the pixel-levelconstraint of a to-be-reconstructed face image through a convolutional neural network, so as to obtain the overall structure information of the face image; selecting a plurality of to-be-repaired face areas from the face image overall structure information by utilizing deep reinforcement learning; gradually repairing each to-be-repaired face region in the plurality of to-be-repaired face regionsthrough the enhanced network to obtain a plurality of repaired face regions; and through the face recognition network and the bidirectional consistency network, carrying out constraint of face category semantic information and face appearance information on the plurality of repaired face regions to obtain a face reconstruction result of the to-be-reconstructed face image. According to the method,deep reinforcement learning is utilized, so that the generated high-definition face recovers rich appearance information, semantic information of the face is reserved, and face super-resolution performance and accuracy can be effectively improved.

Description

technical field [0001] The present invention relates to the technical field of computer vision, in particular to a face super-resolution method and device for semantic reconstruction based on deep reinforcement learning. Background technique [0002] In recent years, face super-resolution has received extensive attention in the field of computer vision. Face super-resolution aims to reconstruct low-resolution face images into high-definition images, and plays an important role in further tasks such as face detection, face alignment and face recognition. Usually, low-resolution face images contain very little information, and face super-resolution needs to use limited information to restore high-definition faces and restore as many face features as possible. Although there are many research works in the field of face super-resolution, it is still a very challenging task to develop an algorithm that simultaneously reconstructs the apparent and semantic information of the face...

AI Technical Summary

Problems solved by technology

Although the existing methods use prior information to play an important role in the realization of face super-resolution, but these information are only auxiliary generated high-definition faces that are closer to the real image in terms of apparent information.

Images