Subband decomposition wherein each subimage contains information about a specific range of scales and orientations provides an important vehicle for discriminating between objects with coherent structure and the random patterns of background noise. We use the cortex transform to create a 32-element feature space representing four scales and eight orientations. Simple hypothesis testing based on variances of the feature elements provides an initial indication of the presence and location of structure whose scale and orientation make it stand out from the background. The feature vectors of these somewhat obvious object pixels are then used in a novel training scheme to derive a linear classifier for subsequent recognition of faint, low-contrast objects. The method is used to detect calcification structure in digitized mammograms, and the results are compared with those of a previously published wavelet technique. Measurable improvements in detector output suggest that the new approach has potential.