Robust face alignment method based on multistage model

Abstract

The invention discloses a robust face alignment method based on a multistage model, and the method comprises the steps: firstly solving an initialization problem caused by a face detector through employing STNs (ASTN) based on adversarial learning, such as rotation and scale changes, so as to obtain a better face boundary frame for face alignment; obtaining initial positions of the face features and corresponding scores of the initial positions by using an hourglass network. Furthermore, the invention also provides a shape dictionary based on the sample, which aims at finding out the low-scorefeature points according to the feature points with high scores, and remarkably improving the human face feature dislocation caused by shielding or background chaos by combining with the face shape constraint.

Description

technical field [0001] The invention relates to face alignment technology, in particular to a robust face alignment method based on a multi-level model. Background technique [0002] Face alignment or facial landmark detection aims to determine a set of predefined human facial landmarks, such as eye corners, eyebrows, and nose tip. Face alignment is an important foundation for advanced vision tasks, such as: face recognition, expression recognition, facial animation, and 3D face modeling. Although great progress has been made on these tasks, face alignment is still challenging due to large-view face variations, lighting conditions, complex expressions, and partial occlusions. [0003] Recently, convolutional neural networks (CNNs) based on heatmap regression have achieved remarkable progress. Hourglass network is a popular human pose estimation method that uses repeated downsampling and upsampling modules to extract features at multiple scales, stacked hourglass network an...

AI Technical Summary

Problems solved by technology

However, even state-of-the-art algorithms cannot correctly locate feature points under extreme occlusion

Worse, inaccurate labeling of feature points leads to inaccurate prediction of affine transformation parameters

[0006] In short, existing face alignment models all have corresponding defects

Images