Semantic mapping of Hindi text-to-image generation using CUB dataset

Area of Science:

• Artificial Intelligence
• Computer Vision
• Natural Language Processing

Background:

• Generative Learning models create new data by learning patterns from existing datasets.
• Text-to-Image (T2I) generation models create images from textual descriptions.
• A significant gap exists in T2I models for regional languages, limiting visual content creation.

Purpose of the Study:

• To propose a semantic mapping for Hindi Text-to-Image generation using Generative Adversarial Networks (GANs).
• To develop and evaluate a regional T2I model trained on Hindi language data.
• To address the limitations of current T2I models in handling regional language inputs.

Main Methods:

• Utilized Generative Adversarial Networks (GANs) for semantic mapping.
• Prepared and pre-processed a Hindi language dataset for model training.
• Employed the Caltech-UCSD Birds 200 (CUB) dataset as a primary data source for experiments.

Main Results:

• The proposed GAN model demonstrated robust performance in generating images from Hindi text.
• Achieved an Inception score of 4.65 and an FID score of 37.17.
• Established a novel semantic alignment between Hindi text and images with an R-precision score of 75.12.

Conclusions:

• The developed model significantly enhances T2I generation capabilities for the Hindi language.
• The research successfully bridges the gap in regional language T2I, enabling realistic visual creation from regional text.
• This work represents a pioneering step in semantic alignment for Hindi text-image generation.