TextCycleGAN FY19
Abstract
There has been much success with image captioning learned on large datasets, but the problem becomes more complex with smaller datasets. Generative adversarial networks (GANs) have shown promise in learning to generalize mappings from smaller datasets. GANs utilize the competition between a discriminator and a generator to strengthen generation. The discriminator identifies whether an input originated from the training set. The generator tries to create an output that the discriminator will falsely detect as being from the training set. With TextCycleGAN, we utilize cycle-consistent GANs (CycleGANs). CycleGANs utilize multiple GANs to learn a mapping between two domains. With CycleGAN, we can improve image captioning performance on smaller datasets by using both the GAN architecture for better generalization and by learning both translations from image to text and text to image. This will be a low-cost software package of a trained CycleGAN model for image captioning to be applied to naval applications.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 01, 2022
- Accession Number
- AD1161684
Entities
People
- Anthony C. Delgado
- Iryna Dzieciuch
- Maurice R. Ayache
- Mitch C. Manzanares
- Mohammad R. Alam
- Nicole A. Isoda
Organizations
- Naval Information Warfare Center Pacific