Face image beautification detection method and system based on optical flow estimation

Abstract

The invention provides a face image beautification detection method and system based on optical flow estimation, and the method comprises the steps: cutting a to-be-processed face image to obtain a face region image, carrying out the deformation processing of the face region image to obtain a face deformation image, and inputting the face region image and the face deformation image into an optical flow extraction network; respectively extracting features of the face region image and the face deformation image for fusion, and performing optical flow extraction on the images after feature fusion by using a deconvolution network to obtain an optical flow field image; downsampling a to-be-processed face image, inputting the downsampled image and the to-be-processed image into the optical flow field detection network, carrying out two-layer cascade feature fusion to obtain a first feature map, a second feature map and a third feature map, respectively calculating common differences between the three feature maps and the optical flow field image, and reversely propagating the common differences to the optical flow field detection network, and the optical flow field detection network converges. Image restoration is carried out according to the optical flow field, and good performance can be shown on face deformation.

Description

technical field [0001] The present application relates to the technical field of computer vision, in particular to a method and system for face image beautification detection based on optical flow estimation. Background technique [0002] For the face image processed by face beautification, some existing technologies can obtain relatively satisfactory results, such as: using image stitching and pattern noise-based detection methods to detect face images after skin grinding, or using color space distribution way to detect images processed by different filters. In face deformation detection, direct use of classification and object detection convolutional neural networks does not work well, but according to the paper "Detecting Photoshopped Faces by Scripting Photoshop", using illumination estimation to model face liquefaction patterns can greatly improve Improve the detection accuracy of liquefied face images. [0003] However, at present, it is not possible to use a unified...

AI Technical Summary

Problems solved by technology

[0003] However, at present, it is not possible to use a unified method for face color beautification detection. If each beautification method needs to use one method for detection, it will cause bloated system; at the same time, the detection method described in the above paper is only applicable to face liquefaction The face deformation is not suitable for other deformation methods, especially the deformation face image based on the moving least squares method

In the face liquefaction detection method described in the above paper, the open source segmentation network expansion residual network DRN is used to estimate the face deformation optical flow of a single image. Since the DRN is not the best in terms of segmentation performance, the subsequent detection method and can not achieve the best performance

Images