Image Generation Using Sketch and Attributes

Abstract

Creating a photographic face picture from a sketch image or written description has always been a difficult problem in computer vision. Sketches normally just provide basic profile information and do not include facial detail. As a result, obtaining exact face characteristics is difficult. To address this issue, we present a photograph translation network that uses characteristics and the generative adversarial network (GAN) ( Image Quality Assessment: From Error Visibility to Structural Similarity Zhou Wang, Member, IEEE, Alan C. Bovik, Fellow, IEEE Hamid R. Sheikh, Student Member, IEEE, and Eero P. Simoncelli, Senior Member, IEEE, apr 2004.). By augmenting the sketch picture with the additional facial attribute feature, it will significantly contribute to the authenticity of the created face. The generator network is made up of two networks: a feature extraction network and a downsampling upsampling network, both of which employ skip connections to reduce the number of layers while maintaining network speed. The purpose of the human network is to check if the created faces have the necessary properties. A photographic face may be created with just a drawing and a simple language description. We have discovered via preliminary testing that sketch drawings include a lot of profile data, whereas attribute vectors provide high-level linguistics data like texture details and colors. When the two datasets are combined, the sketch gives approximately defined data while the attribute provides natural texture information, allowing you to create realistic photographic faces.