2023_programme: The performance of the multi-look coherence technique for target detection

• Session: 15. Towards Automatic Target Recognition. Detection, Classification and Modelling
  Organiser(s): Johannes Groen, Yan Pailhas, Roy Edgar Hansen, Jessica Topple and Narada Warakagoda
• Lecture: The performance of the multi-look coherence technique for target detection [invited]
  Paper ID: 1875
  Author(s): Lyons Anthony, Williams David
  Presenter: Lyons Anthony
  Abstract: Multi-look coherence explores the information content of images by splitting the total spectral bandwidth of a complex synthetic aperture sonar (SAS) image into sub-bands in the wavenumber domain, reforming those sub-bands into multiple complex images, and then calculating the coherence between these multiple looks. The magnitude of the complex correlation coefficient between pixels from different sub-looks may be used as a detection metric as these pixels are expected to be high over a large range of look angles for point scatterers and high for a smaller range of angles for scattering from smooth surfaces. Therefore, spectral coherence may be optimally suited for separating man-made targets from the interfering background clutter as the targets will have features that scatter coherently versus the random seafloor which will scatter incoherently. Although initial small-scale rail-SAS studies on the multi-look coherence technique have demonstrated the potential of the multi-look coherence technique for improved target detection and clutter reduction, the work presented here will give a first look at performance estimated using a large number of images collected over a variety of seafloor environments. ROC curves for multilook coherence detection formed using data from this larger dataset should yield results closer to those which might be expected when using this technique in an operational setting.\n\nMid-frequency (10’s of kHz center frequency) datasets from an experimental SAS system were used for the initial look at detection performance of the multilook coherence technique. A variety of strategies were used to reduce false detections, including averaging pixels to exclude small ‘coherence clutter’ objects, using images formed from only the lower or upper frequency band, and using non-adjacent looks to try to remove objects that had coherence ‘glints’ as opposed to broader angular coherence. All the various strategies improved performance with respect to reducing false alarms, with the best results including all strategies combined (i.e., using only the upper half of the frequency band, followed by using separated looks, followed by averaging pixels in the coherence maps). Performance for the multi-look coherence technique was similar to that of an intensity-based detector for some of the seafloor areas, but markedly worse for others. Seafloors with larger numbers of small compact objects or rocky facets were areas of high clutter in terms of coherence. Performance may be increased by using the multilook coherence technique in conjunction with the more familiar intensity-based detection methods.
• Corresponding author: Dr Anthony Lyons
  Affiliation: University of New Hampshire
  Country: United States
  e-mail: