Three-dimensional face model-based face key point detection method

Abstract

The present invention relates to a three-dimensional face model-based face key point detection method. The method includes the following steps that: step 01, a face image and the initial parameter of a three-dimensional face model are obtained from a face training sample; step 02, a posture adaptive feature and a normalized coordinate code are generated according to the face image and the initial parameter; step 03, transformation and fusion are performed on the posture adaptive feature and the normalized coordinate code through using a convolutional neural network, so that a real residual and the parameter residual of the initial parameter are obtained; step 04, the initial parameter is updated according to the parameter residual, and the method shifts to the step 02 until the parameter residual reaches a preset threshold; and step 05, the three-dimensional face model is updated through using the parameter residual which reaches the preset threshold, and key points on the three-dimensional face model are acquired. With the method of the present invention adopted, face key point detection under a full-posture condition is realized.

Description

technical field [0001] The invention belongs to the technical field of image processing and pattern recognition, and in particular relates to a face key point detection method based on a three-dimensional face model. Background technique [0002] The key points of the face are a series of points with fixed semantics on the face, such as the corners of the eyes, the tip of the nose, and the corners of the mouth. In computer vision based on face understanding, detecting key points is an important preprocessing step. Most face analysis systems need to perform key point detection first, so as to have an accurate understanding of the distribution of facial features of the face, so as to extract features at the specified position of the face. However, most current key point detection methods can only deal with faces below medium poses, that is, faces with a deflection angle (yaw) less than 45°. Face key point detection under large poses (yaw angles can reach 90°) has always been a...

AI Technical Summary

Problems solved by technology

However, this technology has the following defects: First, this technology uses Euler angles to describe the rotation of the face, and Euler angles will cause ambiguity due to gimbal deadlock in large poses; second, this technology only uses the image perspective The input features of the original image are directly sent to the convolutional neural network, and the intermediate result image can be used for gradual correction in the cascade, so as to further improve the fitting accuracy; finally, this technology does not modify the model when training the convolutional neural network. The priority of parameters is effectively modeled, so that the fitting performance of convolutional neural networks is spread over some secondary parameters.

Images