Neural architecture of 3D face modelling using generative adversarial networks

• Published in: AIP conference proceedings Vol. 2707; no. 1
• Main Authors: Kasyap, Varanasi L. V. S. K. B., Bhagavan, V. Srinivasa
• Format: Journal Article Conference Proceeding
• Language: English
• Published: Melville American Institute of Physics 09.05.2023
• Subjects: Artificial neural networks; Feature maps; Generative adversarial networks; Image classification; Image processing; Modelling; Neural networks; Optimization; Recurrent neural networks; Spatial data; Three dimensional models; Training
• ISSN: 0094-243X; 1551-7616
• DOI: 10.1063/5.0143020

3D Face modelling is not same as 2D Face image generation using DeepFake. This paper suggests a model, in solving the problem of responsive 3D face generation using less training data. By using Deep Convolutional Neural Networks (CNNs), the loss function is defined on feature maps. Optimization problem is solved using Stochastic Gradient Descent (SGD). Generative Adversarial Networks (GANs) are used here to generate 3D Face Model from feature maps. The key contribution of work is finding the regional area in the given face spatial data by coalescence of two techniques and adding into feature vector. Emotional synthesizer is also proposed in the model, to make 3D face realistic by scrambling emotions on 3D Face model. Features are extracted from input data (video clips, images) using CNN and used in training Recurrent Neural Network (RNN) makes it to classify the image to be progressed or not. This model is evaluated against dataset generated with 30 people in laboratory and validates the acceptable performance and boosts up the Inception Score (IS) in 3D Face generation with contemplate limits.