Deep Transductive Transfer Learning for Automatic Target Classification: Visible to MWIR/LWIR
Abstract
In this project, we address the problem of domain translation from a labeled source domain to an unlabeled target domain. we assume that the source dataset is in the Visible (VIS) spectrum where a pre-trained classifier (called the optimal VIS classifier) trained on a large-scale VIS dataset is available, and the target dataset is in the Infrared (IR) domain which is not labeled where a classifier needs to be designed. Our proposed framework is called a task-driven deep transductive transfer learning since the task is to design a classifier in the target domain using the available labeled source domain dataset, the optimal source classifier and the unlabeled target dataset. We assume unpaired source and target training datasets and propose an image-to-image domain translation algorithm based on cycleGAN framework. We construct a classifier in the target domain using the a pre-trained source classifier. We formulate a new objective function where we integrate the loss terms from the source and target classifiers into the conventional cycleGAN adversarial loss to train our domain translation model as well as the target classifier. We will demonstrate our transductive transfer learning model or Automatic Target Recognition (ATR) on two different datasets; the dual spectrum SENSIAC dataset containing a large collection of visible and mid-wave infrared (MWIR) imagery collected by the US Army Night Vision and Electronic Sensors Directorate (NVESD) and the Multi-Domain Smart Sensors (MDSS) dataset from the US NVESD containing a large collection of paired MW and LW infrared images with military vehicles. Our goal is to demonstrate a cross-domain transfer learning model to synthesize images of military vehicles from source to target domain (or vice versa) and at the same time construct a classifier/detector for the target domain using the synthesized target images. We will evaluate the quality of our image-to-image domain adaption model based on the quality of the reconstructed images and the performance of the target classifier based on the ROC plots.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 14, 2022
- Source ID
- W911NF2210117
Entities
People
- Nasser M. Nasrabadi
Organizations
- Army Contracting Command
- United States Army
- West Virginia University