Automated Synthetic Hyperspectral Image Generation for Clutter Complexity Metric Development

摘要

Imaging sensors and automatic target recognition (ATR) algorithms are an integral part of modern combat systems. We present a method to automate the efficient synthesis of hyperspectral images used as aid in the evaluation and development of ATR algorithms. To ensure reliable inferences from these processes, it is required that the different levels of difficulty for ATR performance are adequately represented in the generated images. We employ the Digital Imaging and Remote Sensing Image Generation (DIRSIG) software for the image synthesis, and model each image as a function of the input parameters needed for the image synthesis. The computational complexity of image generation makes gradient-based, and similar adaptive schemes inappropriate for sampling this multidimensional function. We present a progressive adaptive sampling algorithm based on the equalization of the histogram of the already obtained samples. The algorithm requires no prior knowledge of how the images vary with the inputs used in their synthesis, and the computational overhead is minimal. The images generated with the aid of this algorithm are compared to those generated from a combination of random, and even spaced input parameters to DIRSIG. An improvement in diversity with respect to ATR performance is recorded for the images generated using the adaptive sampling algorithm.