Sketch guided face synthesis using conditional variational autoencoder

Edson Masao Odake Junior, Eduardo Parente Ribeiro

DOI: 10.14209/sbrt.2024.1571027732

Evento: XLII Simpósio Brasileiro de Telecomunicações e Processamento de Sinais (SBrT2024)

Keywords: Conditional Variational Autoencoder Photorealistic Face Synthesis Forensic Sketch Image-to-Image Translation

Abstract

Forensic sketches, often the only visual leads in criminal cases, typically lack the detail and realism that can help in public identification tasks. This paper presents a Conditional Variational Autoencoder (CVAE) approach for transforming forensic sketches into photorealistic facial images. Using a stochastic edge map extraction of images from a dataset, our model bypasses the need for manually paired sketch-photo databases, enhancing scalability. The model was evaluated on several metrics, demonstrating the capability of working on different image styles.

Download