Text-To-Image Synthesis Using KTGAN

Abstract

This paper proposes a new model, Knowledge Transfer Generative Adversarial Network (KTGAN), for text-to-image synthesis.  A couple of new methods are used such as an Alternate Attention Transfer Mechanism (AATM) and a Semantic Distillation Mechanism (SDM), to assist generator higher connect inter functional hole in linking textual content and picture. AATM modifies phrase interest weights and interest weights of picture segment one after other, to gradually spotlight necessary phrase statistics and enrich small print of generated images. The semantic distillation mechanism makes use of picture encoder skilled in the Image-to-Image project to information coaching of textual content encoder in Text-to-Image process, to produce higher textual content points and greater excellent pictures. By significant investigational testing on two public data sets, KTGAN surpass the existing approach notably, and additionally attains the comparative effects over one-of-a-kind comparison metrics.