Özet:

Abstract In order to create photographic images from semantic text descriptions with high resolution and visual integrity, text-to-image synthesis is used. But if the resolution of the images is increased, network complexity and processing needs become more difficult. To generate high-resolution images, existing models use huge parameter networks and computationally expensive methods, which leads to unstable training processes and expensive training. In this research, we propose a novel semi-supervised image generation model (SIGTI) that uses both labelled and unlabeled datasets for text-to-image conversion. The labelled dataset, which contains text and associated images, is used throughout the training phase. From the text data, we extract features including N-grams, better TF-IDFs, and BOWs. These features are then used to train the feature set of the NIC semi-supervised image creation model, which combines an upgraded GAN and Deep CNN. The unlabeled dataset with simply text is used during the testing phase. In order to create the appropriate relevant image, we extract the N-gram, enhanced TF-IDF, and BOW features from the text and compare them with the trained features using the proposed NIC model. To determine how well our suggested model works, we evaluate it thoroughly and compare its performance to other established methods.

Anahtar Kelimeler: