Generating Sketch Faces from Text Descriptions: Dataset and Algorithm

Sketching, as a form of painting, is widely used to pursue fugitives. The main method of modern criminal investigation sketching is for artists to sketch suspects' faces based on witness descriptions. However, to ensure that the sketches match the witness descriptions, the artists must possess both professional expertise and a substantial reservoir of experience. The process is inherently time-consuming. As a consequence of developments in the field of generative adversarial networks, it is possible to quickly generate corresponding sketch face images from text descriptions. However, this research is challenging due to two facts: first, there is no available dataset of face attribute text to sketches; second, due to the abstractness and simplicity of sketch images, it is difficult to synthesize high-quality face sketches. In this study, our contributions are twofold: 1) We propose a novel text-to-sketch face network, MDPGAN. 2) We build a comprehensive dataset, CelebASketch-HQ, which contains high-quality face sketch images, each accompanied by ten corresponding text descriptions. Finally, experimental evidence substantiates the efficacy of our approach.