Text-to-image generation with enhanced GANs: Bridging semantic gaps using RNN and CNN.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Text-to-image generation aims to create realistic images from textual descriptions.
• Achieving consistent realism in generated images remains a significant challenge.

Purpose of the Study:

• To propose a neural network-based model for generating high-quality images from text.
• To evaluate the model's performance on standard and modified datasets.

Main Methods:

• A Generative Adversarial Network (GAN) framework was employed, integrating Recurrent Neural Networks (RNNs) for text processing and Convolutional Neural Networks (CNNs) for image feature extraction.
• The model was trained on the Oxford 102-flowers dataset and a modified version, Oxford 102 flowers (beta), with resolutions up to 256x256.
• Generator and discriminator losses were calculated, and performance was assessed using Inception Score and Peak Signal-to-Noise Ratio (PSNR).

Main Results:

• The model achieved an Inception Score of 4.15 (64x64 resolution) on the original dataset and 3.97 (128x128 resolution) on the modified dataset.
• PSNR values of 28.25 dB and 30.12 dB were recorded on the original and annotated datasets, respectively.
• The proposed methodology demonstrated superior performance compared to existing models.

Conclusions:

• The developed neural network model effectively generates realistic images from text descriptions.
• The model shows promising results on both standard and augmented datasets, outperforming existing approaches.
• Further research can explore higher resolutions and diverse datasets.