Deflection face correction method based on generative adversarial network improved structure

Abstract

The invention discloses a deflection face correction method based on an improved structure of a generative adversarial network, and the method needs a plurality of deflection faces and face pairs formed by corresponding front faces during training, and only needs to deflect the faces during testing. The method comprises the following processing steps: (1) carrying out feature point detection on ahuman face, and extracting human face region blocks (eyes, nose and mouth) with fixed sizes; (2) respectively inputting the face region blocks and the overall face in the step (1) into a local channeland an overall channel to obtain a result after the local channel and the overall channel are corrected; (3) fusing the output of the local channel and the output of the overall channel, and settingthe pixel of the overlapping region as the maximum value on the region to obtain a final generated face; (4) inputting the generated face and the front face into a discriminator and a classifier to ensure that the accuracy and identity of the generated face are consistent; and (5) storing the network model obtained by training for testing. The adopted BEGAN network is simple in structure and efficient, and the accuracy and speed of deflection face correction are improved to a certain extent.

Description

technical field [0001] The invention relates to a deflected human face correction method based on an improved structure of a generative confrontation network, which belongs to the field of computer vision. Background technique [0002] With the continuous development of deep learning, many breakthroughs have been made in the research of face recognition. The recognition algorithm based on deep learning even exceeds the level of human naked eyes. However, most of these studies are based on the premise of frontal or approximate frontal faces, so these studies have certain limitations. Evidence shows that even the best-performing frontal face recognition methods suffer a significant drop in recognition rates at large angles of deflection. For face recognition under pose changes, existing methods can be roughly divided into the following three categories: feature extraction methods based on pose robustness, methods based on frontal face generation, and methods based on subspace...

AI Technical Summary

Problems solved by technology

[0006] Purpose of the invention: In view of the above problems, in the TP-GAN method of two-dimensional face conversion method, there are many training difficulties and the problem of easy mode collapse for the generation confrontation network DCGAN structure adopted by TP-GAN, and its impact on human The multi-scale image of the face is collected, which makes the training process relatively complicated. The present invention proposes a deflected face conversion method based on the improved structure of the generative confrontation network.

Images