Accession Number:

AD1168651

Title:

TextCycleGan FY20

Descriptive Note:

[Technical Report, Final Report]

Corporate Author:

NAVAL UNDERSEA WARFARE CENTER SAN DIEGO CA

Personal Author(s):

• Alam, Mohammad R.
• Isoda, Nicole A.
• Manzanares, Mitch C.
• Delgado, Anthony C.
• Panggabean, Antonius F.

Report Date:

2022-05-01

Pagination or Media Count:

32

Abstract:

In this report, we discuss improvements to TextCycleGAN a cycle-consistent generative adversarial network CycleGAN for image captioning. CycleGANs train separate Generative Adversarial Networks GANs to learn mappings between multiple domains and strengthens each individual mapping with cycle consistency loss. As such, with CycleGANs we can create a better image captioning generator by jointly training an image synthesis generator. Since cycle-consistency ensures minimal change with recreation of the input, this offers a unique challenge for image captioning due to the many-to-many nature of the mapping from images to captions and vice-versa. We will further discuss how we tackle this many-to-many challenge as well as both image captioning and image synthesis in the report.

Descriptors:

• natural language processing
• language
• computer vision
• information warfare
• training
• united states government
• governments
• natural languages
• united states
• automated text summarization
• consistency
• generators
• information operations
• military research
• probability
• abstracts
• computer languages

Subject Categories:

• Computer Programming and Software
• Cybernetics

Distribution Statement:

[A, Approved For Public Release]