An optical character reader is characterized in that the reader determines accurately the range occupied by the major portion of each character in each character line. This is accomplished by dividing an image read from a document into character liens and counting the number of picture elements in series with each picture element forming each character on each character line in each direction of a plurality of predetermined directions. Each picture element is provided with a code data indicating the direction having the largest number of picture elements having a character. If the picture elements between two picture elements allotted with a specific code data for the direction along each character line do not have the specific data code, the original code data is replaced with the specific code data. The number of picture elements allotted with the specific code data for each character line are then counted and the position with the largest number of picture elements having a character is detected.

A histogram is prepared using the heights of characters in an input alphanumeric character image data set. Based on this histogram, reference line characters are detected from the character row. The upper and lower extracting ordinates of each character in the character row are obtained. Then, the average of each of the upper and lower extracting ordinates in the character row is determined as the ordinate of a reference line for the character row. Thus obtained reference line is used for character recognition.

A model for a previously defined form is provided in terms of directional features for particular evaluation points and shift vectors from one evaluation point to the next. An input image is represented by density gradients for different directional planes. Shift operations for a counter plane which is used to accumulate values for evaluating the similarity between the image and the model, and add operations in which values from input directional planes as specified by the model are added to the counter plane, are performed repeatedly by parallel processing to find the total match with the model for the form.

The present invention relates to a method for determining an orientation of a line segment in a contour in a local area of a binary contour image. The number of pixels having a predetermined value and located in each orientation in the local area, is obtained, and it is determined that a line segment exists in an orientation in which the number of pixels is large. Further, when it is determined that the number of pixels located in each of a plurality of orientations and having a predetermined value is greater than a sum of the numbers of pixels in orientations adjacent to said each orientation on both sides thereof, it is determined that a line segment exists in said each orientation. Two orientations in which large numbers of pixels are located are detected, and it is determined whether the point is a constituent of a line segment, a line segment of the sub-straight-line form, or a near-corner point, depending on angles made by the two orientations is equal to 180.degree., or a difference of the angles made by the two orientations from 180.degree.. Contiguously arrayed pixels in the same orientation are detected as a line segment. A position and an orientation are obtained from points near a corner, and representative values of the positions and orientations of contiguously arrayed near-corner points are determined.

A character recognizer includes a memory for storing low resolution data of a group of characters to be recognized in an image and a memory for storing high resolution data of the group of characters. Wide range image data for the characters stored in the high resolution memory is selected on the basis of narrower range image data for corresponding characters stored in the low resolution memory. Character recognition is performed according to the selected character image data.