Fast Learning VIEWNET Architectures for Recognizing 3-D Objects from Multiple 2-D Views

The recognition of 3-D objects from sequences of their 2-D views is modeled by a family of self-organizing neural architectures, called VIEWNET, that use View Information Encoded With NETworks. VIEWNET incorporates a preprocessor that generates a compressed but 2-D invariant representation of an image, a supervised incremental learning system that classifies the preprocessed representations into 2-D view categories whose outputs arc combined into 3-D invariant object categories, and a working memory that makes a 3-D object prediction by a.ccurnulating evidence from 3-D object category nodes as multiple 2-D views are experienced. 'I'hc simplest VIEWNE'I' achieves high recognition scores without the need to explicitly code the tcmpora.l order of 2-D views in working memory. Working memories are also discussed that save memory resources by implicitly coding temporal order in terms of the relative activity of 2-D view category nodes, rather than as explicit 2-D view transitions. Varia.nts of the VIEWNE'I' architecture may also be used for scene understanding by using a preprocessor and classifier that can determine both What objects arc in a scene and Where they arc located. 'I'hc present VIEWNE'f preprocessor includes the COHT-X 2 filter, which discounts the illurnina.nt, regularizes and completes figural boundaries, and suppresseo image noise. 'I'his boundary segmentation is rendered invariant. under 2-D t.ra.nsla.t.ion, rotation, and dilation by use of a log-polar transform. The invariant spectra undergo Gaussian coarse coding to further reduce noise and 3-D foreshortening effects, and to increase generalization. 'I'hcsc compressed cocleo are input into the classifier, a supervised learning system baocd on the fuzzy AJ(I'iVIAP algorithm. Fuzzy AltfMAP learns 2-D view categories that are invariant under 2-D image translation, rotation, and dilation a.s well as 3-D image tra.noformations that do not cause a. predictive error. Evidence from scquenceo of 2-D view categories converges at :J-D object nodes that generate a responoc invariant under changes of 2-D view. 'I'hese 3-D object nodes input to a working mernory that accumulateo evidence over time to irnprove object recognition. ln the oimplest working rnernory, each occurrence (nonoccurrence) of a. 2-D view category increa.oeo (decreases) the correoponding nocle'0 activity in working rncrnory, 'I'he rna.xirnally active node is used to predict the il-D object. Recognition is studied with noisy and clean ima.geo using slow ancl fast learning. Slow learning at the fuzzy AH'I'MAP rnap field is adapted to learn the conditional probability of the ~l-D object given the selected 2D view category. VIEWNET is clernonstrated on an MJ'I' Lincoln Laboratory database of l28x128 2-D views of aircraft with and without additive noise. i\ recognition rate of up to 90% is achieved with one 2-D view and of up to 98.5% correct with three 2-D views. 'I'hc properties of 2-D view and 3-D object category nodes are compared with those of cello in monkey inferotcmporal cortex.