Extracting information-dense vectors from images for neural network classifiers

Summary form only given. Neural network classifiers are being developed to score biomedical antigen/antibody reactions used in blood typing. Images of the reactions are captured and digitized with a high-resolution charge-coupled device camera using microscope optics. These 262, 144 element image arrays are too large for direct input to present-day backpropagation neural networks. Digital sampling detector architectures have been synthesized which are applied to the two-dimensional Fourier transform of the original image, yielding dense vectors several orders of magnitude smaller than the input image. These vectors are then applied to multiple three- and four-layer backpropagation networks. A typical network has an input layer of 48-280 elements, hidden layers of 12-64 elements, and an output layer of 8-20 elements. The multiple neural nets classify the reactions into categories with associated confidence scores and check for the presence and extent of a variety of flaw and error conditions such as poor focus, inadequate sample or reagent, cracks in the plastic tray, and bubbles or shadows in the sample