Infrared Target Recognition Using Realistic Training Images Generated by Modifying Latent Features of an Encoder-Decoder Network

Generating realistic images has been a challenging problem in computer vision, with many researchers focusing on novel methods and datasets to produce benchmark results. Our motivation for the same arises from the dearth of real training images for recognizing targets in infrared imagery. We propose an encoder-decoder architecture for generating realistic medium wave infrared images of targets at various azimuth angles, in day or night conditions, and at different ranges. Specifically, we use a CNN (Convolutional Neural Network)-based siamese autoencoder network that modifies the latent space embedding of a given input view to produce a novel output view. First, we train this network with a limited set of real images of the targets, and show that it can generate new and previously unseen views of the same. We show that the network operates in the nonlinear feature subspace and learns the underlying manifold to develop a semantic understanding of the targets. We use the structural similarity index measure to quantify how the generated and real images of targets compare. Finally, we show classifiers trained with the generated images are able to recognize targets in real test images.