JPH0245230B2 - - Google Patents

Description

TECHNICAL FIELD The present invention relates to a fast and simple feature normalization circuit.

(Background technology) Conventionally, as a character recognition method, for example, "Introduction to Character Recognition" edited by Shinichiro Hashimoto (published March 20, 1980)
Published in Telecommunications Association, pages 110-113
There are recognition methods such as Glucksman's method or topological structuring method.

Glucksman's method, as shown in Figure 2,
This method analyzes the situation under which each point on a white background in a character pattern area exists by observing the character pattern in the vertical, horizontal, and vertical directions from that point. In other words, when looking at a character by extending a straight line from a certain point in the vertical and horizontal directions, ternary encoding is performed in which the character is defined as 0 if it does not collide with the character line, 1 if it collides once, and 2 if it collides two or more times. For example, point P in Figure 2 can be obtained using this method.
(2012). However, the directions are shown in the order of left, top, right.

The code assigned to each of these points is called a feature.
In Glucksman's method, identification processing is performed based on the type of each code (that is, each feature) and the area of the region (in other words, the result of counting the number of appearances of each feature).

Furthermore, in the case of character recognition using the topological structuring method, identification processing is performed based on the type and number of higher-order features present in the input character pattern.

In the character recognition method described above, after counting the number of occurrences of each feature in a feature pattern register that stores feature patterns extracted from character/figure patterns, changes in character size, especially handwritten characters, are calculated. In order to absorb differences in size depending on the writer, the counting results are often normalized based on the size of the characters. Equation (1) is an example of the normalization equation.

F _i =f _i ×C/Δx×Δy (1) However, f _i is the result of counting the number of appearances of each feature i, Δx
and Δy are the horizontal and vertical lengths of the character circumscribing frame, C is an integerization constant, and F _i is the normalization result of feature i. However, as in the example above, normalization requires multiplication and division functions, which results in a complicated circuit when converted into a device, and multiplication and division after counting the number of occurrences of each feature i. The processing time was reduced due to the calculation.

(Problem of the Invention) The present invention aims to improve the above-mentioned drawbacks of the conventional technology, and its purpose is to provide a feature normalization circuit that is fast and has a simple configuration. The count result f _i of the number of occurrences of each feature i in the feature pattern register that stores the feature pattern created by extracting the features is calculated according to the font size; F _i = f _i ×C/Δx × Δy f _i : Result C of counting the number of occurrences of each feature i; Integerization constants Δx, Δy; Length F _i in the vertical and horizontal directions of the character circumscribing frame; In the feature normalization circuit that normalizes based on the normalization result, Δx
and Δy, a fixed memory that receives the output of the register as an address input and stores C/Δx×Δy at the address, and a memory address corresponding to each feature i and whose initial value is Each time each feature i is detected by scanning the cleared feature memory and the feature pattern register, the content of the memory address corresponding to each feature i in the feature memory and the output of the fixed memory are added; an adder for re-storing the sum in the memory address of the feature memory; and means for repeatedly performing addition by the adder f _i times, which is the result of counting the number of occurrences of each feature i. It is in the feature normalization circuit as shown in FIG.

(Structure and operation of the invention) FIG. 1 is a block diagram of an embodiment in the character recognition device of the invention, in which 1 is a feature number input as an address number, 2 is a character width register, and 3 is a character height register. , 4 is a fixed memory constituted by ROM, 5 is an adder, and 6 is a feature memory. Note that a feature pattern register (not shown) stores a feature pattern created in advance by extracting a plurality of features from an input character/figure pattern.

This example is an example of the above formula (1). Fixed memory 4 (hereinafter referred to as ROM) is a fixed memory whose inputs are character width (Δx) and character height (Δy), and the memory contents specified by addresses by Δx and Δy are expressed by the following equation (2). It will be done.

C/Δx×Δy (2) where C is an integer conversion constant in equation (1), and in this embodiment, C=100, and the result of equation (2), that is, the output of the ROM, is a real number type expressed in binary. (For example, when the value is 0.5 in decimal notation, the output of the ROM is 0.1, and . is a decimal point.) The operation of this embodiment is as follows.

First, the contents of the feature memory 6 are cleared in advance. Next, the size of the character circumscribing frame of the input character pattern is set in the character width register 2 and character height register 3. When feature number i of feature i is input as feature number input 1 to specify the address of feature memory 6, the contents of address i of feature memory 6 are read out, and are read out by character width register 2 and character height register 3. The contents of the ROM specified by the address are added by the adder 5, and the result is written to address i of the feature memory 6 again. The process from inputting the feature number i of each feature i to writing it into the feature memory as described above is performed by using a scanning means (not shown) to scan a feature pattern register that stores feature patterns created by extracting multiple features. It is repeated each time each feature i is detected by scanning,
Ultimately, the process is repeated f _i times, which is the result of counting the number of occurrences of each feature i. The integer part of the final content at address i in the feature memory 6 is the normalized result of feature i shown by equation (1).

As explained above, in this example, equation (1) is transformed into the following equation (3).
This is nothing more than a transformation and execution as in the formula.

F _i =C/Δx×Δy×f _i (3) In other words, in this example, the normalized result is calculated by multiplying C/Δx×Δy (ROM output) by the number of occurrences of feature i. .

As explained above, in this embodiment, calculations can be made only by adding the normalized results of features at the same time as scanning the feature pattern register in the character recognition device without performing multiplication or division. It has the advantage of achieving high-speed processing with a simple circuit.

(Effects of the Invention) Since the present invention simultaneously performs feature counting and normalization by using a fixed memory and an addition circuit, it is possible to realize a character recognition device that is fast and has a simple circuit configuration. can.

[Brief explanation of the drawing]

FIG. 1 is a configuration diagram of an embodiment of the present invention, and FIG.
The figure is a diagram illustrating feature extraction using the conventional Glucksman method. 1...Characteristic number input, 2...Character width register,
3...Character height register, 4...Fixed memory, 5
...Adder, 6...Feature memory.

Claims (1)

[Scope of Claims] 1. The count result f _i of the number of occurrences of each feature i in a feature pattern register that stores a feature pattern created by extracting a plurality of features from a character/figure pattern is calculated according to the size of the character. ; F _i = f _i ×C/Δx × Δy f _i ; Result of counting the number of occurrences of each feature i C; Integer conversion constants Δx, Δy; Length of character circumscribing frame in vertical and horizontal directions F _i ; Normal In the feature normalization circuit that normalizes based on the normalization result, Δx
and Δy, a fixed memory that receives the output of the register as an address input and stores C/Δx×Δy at the address, and a memory address corresponding to each feature i and whose initial value is Each time each feature i is detected by scanning the cleared feature memory and the feature pattern register, the content of the memory address corresponding to each feature i in the feature memory and the output of the fixed memory are added; an adder for re-storing the sum in the memory address of the feature memory; and means for repeatedly performing addition by the adder f _i times, which is the result of counting the number of occurrences of each feature i. A feature normalization circuit comprising: