A Facial Feature Point Localization Method Based on End-to-End Recurrent Network

Abstract

The invention discloses a method for locating facial feature points based on an end-to-end recurrent network, comprising the following steps: (1) data preparation stage, (2) end-to-end model design stage, (3) model training stage, (4) The stage of model testing to obtain the facial feature point location result of the face. In the present invention, the depth feature embedded with shape information is used in conjunction with the cyclic neural network, which greatly improves the positioning accuracy of facial feature points. In addition, the present invention uses inherited neural network features, greatly reduces the amount of model calculation, and can achieve super real-time facial feature point positioning speed under the condition of maintaining better facial feature positioning accuracy.

Description

technical field [0001] The invention relates to a positioning method, in particular to a facial feature point positioning method based on an end-to-end recurrent network. Background technique [0002] The main goal of facial feature point positioning is to automatically locate key feature points of the face based on the input face, such as eyes, nose, mouth, and facial contours. This technology is widely used in face recognition, expression recognition, 3D reconstruction of face, and synthesis of face animation. Most of the current facial feature point localization methods are based on the deep neural network framework. The main methods are: [0003] 1) Coarse-to-fine facial feature point localization based on cascade structure or module division. Related patents: CN105981041A, CN105868769A. Its main technical means are: use multi-level model cascade and perform facial feature point positioning in a coarse-to-fine manner, and refine the position of facial feature points ...

AI Technical Summary

Problems solved by technology

Problems and shortcomings: This type of method is not robust to the facial feature point positioning of large-pose faces. The cascaded or sub-module framework is more sensitive to the given initial facial feature points. Once the initial facial feature point position is far away from the target position , the positioning of the final feature points will have a large deviation from the target; secondly, the cascaded framework is more likely to enter poor local optimization points during the training process, resulting in poor performance of the final model

Problems and disadvantages: Hard classification based on face angle is not necessarily the most suitable classification method for facial feature point positioning tasks, and this hard classification method may not be robust to facial feature point positioning of samples at category boundaries

Images