The researchers at IIIT Allahabad have proposed T2CI-GAN, a novel deep learning model that generates compressed images from input text. It was developed by researchers from the Computer Vision and Biometrics Laboratory at the institute and will form a robust groundwork for future image-storing and content-sharing technologies.
Existing methods of generating images from input texts utilize GANs (generative adversarial networks) for image generation and then compress the generated images in the following step.
This novel method expands the existing ones by directly generating compressed images reducing the workload and processing time.
The researchers created two GAN-based models to generate compressed images. The first was trained using a dataset of compressed JPEG DCT (discrete cosine transform) images, and the second used a set of RGB photos. The second model was developed to enhance the production of JPEG-compressed DCT representations.
T2CI-GAN will be essential as machines need data to be read or understood in compressed forms. The model currently only produces JPEG-compressed images. Therefore, the long-term objective of the researchers is to expand it to produce photos in any compressed form without any limitations on the compression algorithm.
To know more, refer to the research paper, ‘T2CI GAN: Text to Compress Image Generation using Generative Adversarial Network.’