Metric Generative Adversarial Network with Gated Attention Unit for Text-to-Image Translation

The text-to-image translation is a technique that involves generating visuals accurately representing the details described in the text. However, challenges remain where the produced image cannot completely capture the semantics of text description. Here, the target object in the produced image is incomplete and the texture architecture of the target object lacks richness and detail. In this research, the Metric Generative Adversarial Network (MGAN) with Gated Attention Unit (GAU) method is proposed to translate text to image. The encoder-decoder architecture is developed including the gating blocks utilizing a decoupling style phase-aware technique. This technique estimates the phase data and magnitude of clean data in parallel and effectively minimises interaction between the phase and magnitude. The convolution-augmented gated attention unit is utilized for capturing the time and frequency dependencies with less execution complexity to attain good outcomes. The MGAN with GAU obtains an IS value of 4.42 on Oxford-102 flowers dataset and an IS value of 5.16 on the CUB Bird dataset, which is higher when compared with the Recurrent Convolutional Generative Adversarial Network (RC-GAN).