Image sets for the training of image processing systems

  • Authors:

  • Shawn Aldridge;Michael Peterson;Britny Herzog

  • Affiliations:

  • University of Alaska Anchorage, Anchorage, AK, USA;University of Hawaii at Hilo, Hilo, HI, USA;University of Alaska Anchorage, Anchorage, AK, USA

  • Venue:
  • Proceedings of the 12th annual conference companion on Genetic and evolutionary computation
  • Year:
  • 2010

Quantified Score

Hi-index 0.00

Visualization

Abstract

State of the Art image compression algorithms utilize Discrete Wavelet Transforms (DWTs), to losslessly compress raw images for storage and transmission. These techniques form the basis for image formats such as JPEG2000, as well as the storage of medical images, such as ultrasound and CT scans. The transmission of defense and aerospace images, such as those taken by unmanned aerial vehicles (UAVs), and satellites, is also built upon these techniques. Recent research has shown that image compression filters can be optimized through the use of evolutionary algorithms (EAs). The images used to train these optimized EAs must be chosen to reflect the likely applications of the image processing system. This paper presents a set of 50 images, for use as training images for UAV and satellite image processing systems. The image set is taken from a diverse range of satellite images, including airports, cities, and military bases considered representative of likely targets for reconnaissance missions. We first evolve filters using each image, and then identify the most effective training images based on average mean squared error (MSE) improvement over existing wavelets when each filter is applied across the entire set of images. The resulting subset of images will provide a starting point for the evolution of other defense application image pro