Year: 2024
Type: conference-paper
Abstract: Text-to-image synthesis is an advanced computer vision task that aims to generate images given textual descriptions. While conventional methods have demonstrated an impressive capacity to render high-quality images, they frequently encounter challenges in the realm of producing fine-grained images that embody complex semantic details. To address this, we propose a novel that integrates Expert Knowledge into an Adversarial model (abbreviated as EKA). By utilizing the BLIP expert model to guide the discriminator, EKA encourages the generator to produce more realistic and fine-grained images with complex semantics. In addition, we design a semantic enhancement learning (SEL) module, which generates the fine-grained semantics feature representation by fusing features before and after affine transformations. To maintain semantic consistency between generated images and texts, we also introduce a semantic alignment loss. Extensive experiments on public benchmark datasets confirm the superiority of our approach.
Authors: Hongzhen Li, Li Liu, Huaxiang Zhang, Xiaofeng Qu, Yachao He
Institutions: Shandong Normal University, University of Jinan
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Awards: G1645261029, G3117759162, G62463823
OpenAlex ID: W4409829986