Human face pose estimation method and system based on knowledge distillation

Abstract

The invention relates to a human face pose estimation method and system based on knowledge distillation, which can compress a human face pose estimation model by applying a feature distillation method on the premise of ensuring the accuracy, the model parameter quantity after feature distillation is less, the reasoning speed of the human face pose is higher, the resource consumption is reduced, The method solves the problems of high resource consumption and low accuracy of a shallow network when a deep network structure is applied to face pose estimation, improves the pose estimation accuracy of a large-angle, fuzzy and mask-wearing face image according to the current application scene demand, improves the robustness of the model, improves the face pose angle prediction effect of the model, in practical application, a face pose angle prediction result in a complex scene is more accurate, and a prediction effect exceeding a ResNet50 network structure is realized by using a ResNet18 basic network.

Description

technical field [0001] The invention relates to a face pose estimation method based on knowledge distillation, which belongs to the technical field of face image recognition. Background technique [0002] In face alignment, face tracking, face recognition and other research work, face pose angle has an important impact, and pose estimation is an essential part of many face analysis tasks. Face pose estimation refers to the calculation of the specific direction of the face or head in three-dimensional space. Euler angles (Yaw, Pitch, Roll) are a commonly used face pose representation. [0003] At present, the face pose estimation method is mainly divided into two types: one is obtained by calculating the corresponding relationship between the key points of the face and the image to the three-dimensional coordinates, the pose result depends on whether the key point detection is accurate, and depends on the standard head model and camera The parameter matrix; the other is to r...

AI Technical Summary

Problems solved by technology

Due to the large number of model parameters, in the inference stage, more computing resources are occupied, and the inference speed needs to be improved.

[0008] 2) The shallow network model predicts a large error in the attitude angle

Using some relatively simple feature extraction networks will affect the attitude evaluation effect and increase the angle error

[0010] 3) There is a large error in the pose estimation of images with large angle poses, blur and masks

[0011] Most of the training data of face pose use 300W or 300W-LP public data sets, and the pose estimation error for large-angle and blurred images is relatively large, and in the current application scenario, the robustness of face pose estimation results with masks is poor

Images