Synthetic aperture radar processing by a multiple scale neural system for boundary and surface representation

摘要: Abstract A neural network model of boundary segmentation and surface representation is developed to process images containing range data gathered by a synthetic aperture radar (SAR) sensor. The processing are accomplished an improved Boundary Countour System (BCS) Feature (FCS), respectively, that have been derived from analyses perceptual neurobiological data. BCS/FCS makes structures such as motor vehicles, roads, buildings more salient interpretable human observers than they in the original imagery. Early ON cells OFF embedded shunting center-surround models preprocessing lateral geniculate nucleus (LGN). Such compensates for illumination gradients, normalizes input dynamic range, extracts local ratio contrasts. cell outputs combined BCS define oriental filters corticla simple cells. Pooling at overcomes complementary deficiencies each type along concave convex contours, enhances simpl;e sensitivity image edges. Oriented filter rectified sensitive opposite contrast polarities pooled complex output stages short-range spatial competition (or endstopping) orientational among hypercomplex Hypercomplex activate long-range cooperative bipole begin group boundaries. Nonlinear feedback between segments regions cooperatively completing regularizing most favored boundaries while suppressing noise weaker groupings. performed three copies small, medium, large scales, whose subsequent interaction distances covary with size filter. Filling-in multiple representations occurs within FCS scale via boundary-gated diffusion process. Diffusion activated normalized LGN OFFfilling-in domains. restricted defined gating signals corresponding segmentation. filled-in opponent OFFsignals subtracted form double representations. These shown any methods be both contrasts background luminance. scales then added yield final multiple-scale output. perform favorably comparison several other techniques speckle removal.