Importance

Loading Importance Indicators...

Abstract

A neural network architecture is provided for optical character recognition from an input image in which the target character may be rotated in the image plane. The architecture includes hidden units whose inputs receive image information from portions of the image which are rotationally distributed. That is, the local link between input units and the hidden units is adequate for rotation of the character in the image. Therefore, regardless of the orientation of the image, it is right side up, or approximately so, with respect to one of the hidden units. The hidden units have corresponding inputs with corresponding weight factors, i.e., symmetric weight sharing. Thus, regardless of the orientation of the image, one of the hidden units will produce a high output value indicative of an upright character. Alternatively, a single hidden unit has groups of inputs, each group having a corresponding set of weight factors. The groups are coupled to input units for rotationally distributed portions of the image. Therefore, for any orientation of the image corresponding with one of the groups of inputs of the hidden unit, the hidden unit produces the same output value. In an preferred embodiment, feature information such as local contour direction information is provided to the input units. The feature information is provided with respect to slices of the image taken in different directions.

Loading the Abstract Image...

First Claim