Facial Depth and Surface Normal Vector Prediction Method Based on Atrous Convolutional Neural Network

Abstract

The face depth and surface normal vector prediction method based on the hole convolutional neural network, including the steps of training the hole convolutional neural network: S1, building a hole convolutional neural network, including multiple convolutional layers connected in sequence, multiple holes Convolution layer and multiple deconvolution layers, wherein each convolution layer is connected with a normalization operation and an excitation operation; S2, initialize the weight value of the hole convolution neural network; S3, input the training picture into the hole convolution neural network In the network, the atrous convolutional neural network is iteratively trained with the goal of minimizing the cost function; the weight value is updated once every iteration; S4, input the test picture into the atrous convolutional neural network obtained by training, and output the corresponding Face depth map and surface normal vector map; S5. According to the output face depth map and face surface normal vector map, judge whether the prediction accuracy of the trained hollow convolutional neural network meets the requirements: if it meets the requirements, end the training; if it does not meet the return step S3 continues training.

Description

technical field [0001] The invention relates to the fields of computer vision and digital image processing, in particular to a method for predicting human face depth and surface normal vectors based on a hollow convolutional neural network. Background technique [0002] Facial depth prediction is a novel and challenging problem in computer vision. Depth prediction of the face is an important part of understanding the geometric relationship of the whole face. Correspondingly, the geometric relationship of the face can well reflect the organs on the face and the entire facial environment. If a better face depth can be obtained The information will provide great help to the face recognition problem. Similarly, it will also help to build a 3D model of the face, which will help solve the problem of 3D reconstruction of the face. However, predicting the depth information of the entire face from an RGB image of a face is itself a problem with a pathological nature, because there a...

AI Technical Summary

Problems solved by technology

For example, with the continuous increase of the convolutional layer, the gradient disappears and the training cannot continue; the receptive field of the deep network is relatively small, the depth map and surface normal vector map trained are not very accurate, and the image is relatively rough

Images