Disentanglement of Latent Representation for Face Vectorization using Stacked Multi-tasked Learners

Disentanglement of the latent space of facial features like eyes, nose, mouth, and brows, also known as face vector regression, aims to identify well-defined subset subspaces for each attribute to permit manipulation on points within the local subspace for synthesis and editing tasks. A delineated feature space gives greater control over the variation in data and aids in the processes of facial synthesis and facial attribute editing. This study addresses this challenge by enabling uncoupled feature encoding of facial attributes while keeping resource utilization to a minimum. Given that the development of generative models is constrained by extensive memory and resource requirements, this work proposes a simple and novel algorithm for disentangling the latent space of facial attributes from an image. The challenge has been addressed by using a neural architecture comprising multi-tasked convolutional neural networks for shared feature extraction and an ensemble of multi-tasked learners for a disentangled manifold of facial attributes. Additionally, the model also comprises a classification module and a reconstruction module to ensure identity preservation of attributes. The model has been deployed with minimal supervision. The optimality of the disentangled manifold can be observed by reconstructing the individual facial features of the image from this space with an accuracy of 99.8%, 99.9%, and 91.3% for the left eye, right eye, and mouth, respectively.