Criminal Portrait Synthesis from Text-to-Image Using Deep Learning

This is the authors' abstract. We don't add key points for this paper.

The traditional forensic sketches, which are constructed from eyewitness accounts, are very often wrong due to the limitations of memory and approach through subjective interpretation by the artist. A new deep learning paradigm is proposed in this research that leverages the notions of Deep Fusion-GAN and Sentence-BERT for more realistic synthesizing of criminal portraits, rendered solely from textual descriptions. The system then encodes text into semantic embeddings by Sentence-BERT and allows precise semantic correspondence between different attributes described by text and the produced facial features. Upon evaluation on the CelebA dataset, the model results in much more realism and accuracy than other baselines such as AttnGAN and forensic sketches. Experimental findings show that the synthesized images have a close reference to textual inputs and include internal details like hairstyle, facial expressions, and accessories. So, this automated framework will be a robust tool for the law enforcement community in using a fullfledged scalable object for identification of suspects and can drastically minimize dependence on inaccurate manual sketches and speed-up investigation processes.