An optical character recognition device including means to define a scanning area which includes portions of a printed character to be scanned by a television camera, said camera scanning discrete points of said area arranged in spaced vertical paths, and including means to measure the optical density of the character portion encountered at each point and to assign a binary-coded numerical value to each point and to store each of said numerical values in a matrix-like fashion corresponding to the respective point's position in said scanning area. Electronic circuitry is also provided to derive a plurality of identifying characteristics from said numerical values which uniquely describe the scanned character, said identifying characteristics then being compared to the identifying characteristics of a predetermined number of reference characters by application of predetermined comparison criteria to determine which reference character is the closest match to the scanned character, said closest reference character is then outputted, and means are also provided for modifying the identifying characteristics of said reference characters to compensate for type irregularities and typing ribbon wear.

A pattern recognition processing system in which variations in a character, pattern or the like, especially in a handwritten one, are suppressed to extract its invariable characteristics, ensuring accurate recognition of the character, pattern or the like. The pattern to be recognized is divided into circumscribed quadrangular areas and scanned in horizontal and vertical directions to extract reflection segments between the vertical frames of the circumscribed quadrangular areas and particular segments between adjacent ones of pattern strokes. Further, endpoints of these reflection and particular segments are checked in directions across the scanning directions to extract those endpoints which are not blocked by the pattern strokes, and the corresponding positions of the co-ordinates of these endpoints on a figure frame in the directions across the scanning directions are encoded to extract characteristics of contours of the pattern to be recognized.

The present invention relates to a character recognition equipment which is capable of recognizing with high accuracy those characters that have poor contrast and produce variation in printing density such as characters printed on industrial products. The character recognition equipment according to the present invention is composed of an input unit group which processes in image in a plurality of window regions that are set on a character image, a feature unit group which performs calculation on values obtained by adding thresholds to the output values therefrom, and an output unit group which performs calculation on values obtained by multiplying output values from the feature unit group and adding the thresholds to the sum total thereof. Here, the thresholds and the weights can be adjusted automatically with an actual object image.

A handprinted symbol recognition system for identifying through computer techniques free-form, unconstrained handprinting in which accuracy is decoupled from efficiency converts a thin-line (skeletonized) figure output from a preprocessing system, in the form of an array of "black" image points from a raster line sampling, into segment-oriented lists. The segment-oriented lists are filtered and data compressed to obtain a reduced more suitable subset of points approximating the symbol for further processing. The topologic features of the image are extracted from the subset. Short segments that are potential spurs are removed, the image is rotated as necessary and normalized as to size. Each segment and their interrelationships are analyzed to extract geometric features for use in conjunction with the topologic features in a logic tree decision mechanism which uses three modules: (1) pre-recognition ("trash" filter), (2) potential symbol identification, and (3) final quality assurance. The pre-recognition module separates all possible non-meaningful images from those the recognition system has a chance of identifying with minimal misclassification. The symbol identification module uses a decision tree approach in which the most reliable and "rugged" features (relative to real-world scanner data) are placed at the top of the tree. The quality assurance module asks more precise questions about the tentatively identified symbol to determine if it is indeed the particular symbol in question, i.e., the symbol meets a sufficient class membership (SCM) criterion.

The figure reducers of a figure pre-processing device are connected in series. A plurality of previously determined deletion judging patterns are classified into a plurality of groups and the picture signal pattern is compared with the previously determined judging patterns wherein the point focused on is excluded from a group in accordance with patterns rotated a predetermined angle around a point on a figure to be scanned. The classified groups are supplied to the series-connected figure reducers as a deletion judging pattern for deleting a point in each of the figure reducers.