JPS598862B2 - Pattern recognition method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a pattern recognition device, and particularly to a recognition method for an optical character reading device.

In optical character reading devices, the input pattern is identified by matching the input pattern with a standard pattern, or by converting the input pattern into a feature code string and passing it through a sequential logic circuit. Also, there are cases where an input pattern is accepted as two or more standard patterns, and it is not clear which category to judge.

Usually, such patterns are rejected as confusing patterns. In other words, if the similarity with the standard pattern closest to the input pattern is Smx, and the similarity with the next closest standard pattern is Smx, then 5mx≧ε1・s
nx≧ε1 ・・・・・・・・・・・・(1)5mx−
5nx≧ε3 ・・・・・・・・・・・・・・・・・・
(2) However, the input pattern is determined to belong to the same category as the standard pattern m only after the threshold values of ε1 and ε2 are reached, and if this condition is not met, the input pattern is rejected.

However, 5mx-5nx<ε2...
...Even if the condition (3) is satisfied and the pattern is rejected, it is often clearly distinguishable by humans.

For example, regarding "7" bs shown in FIG. 1, humans know that the upper right branch of "7" is short and that of "T" is long, and can distinguish between the two. The present invention is characterized in that branch lengths are used to distinguish such patterns. When comparing branch lengths, it is easier to compare if you know which pattern the input pattern is close to.Therefore, in the judgment section, if the input pattern is similar to two or more standard patterns and is rejected. The information on which standard pattern is closest is also output, and the lengths of the branches are compared based on that information. Hereinafter, specific embodiments of the present invention will be described with reference to the drawings. FIG. 2 is a schematic configuration diagram showing the pattern recognition device of the present invention. Characters written on paper are converted into electrical signals by a photoelectric conversion device and sent to a preprocessing device. The preprocessing device removes noise, normalizes the size, normalizes the position, and cuts out each character. Each character pattern is converted into a line figure with a line width of 1 bit by a narrowing circuit. Feature points such as end points and branch points are extracted from this line figure. Next, a direction code string is created from the line figure by a direction coding circuit, with each feature point as the starting and ending point. A direction encoding circuit represents the direction of a minute line element with a direction code in order from the start end, and encodes it along the line to the end.For each branch, along with the direction code string, Create a table called a branch table containing coordinates, branch lengths, etc. In the above explanation, we took as an example a case in which a human-powered pattern is thinned into a line figure, and feature points, direction codes, and branch lengths are obtained from the line figure, but there are various other possible feature extraction circuits that can be considered as feature extraction circuits. .

It is clear that other feature extraction circuits may be used as long as the length of the branches can be determined by some method. If the input pattern is expressed as a code string for each branch,
Make a correspondence between the standard pattern and each branch and perform mapping.

Cm the category of the standard pattern closest to the input pattern
, the similarity is Smx, the category of the next closest technical standard pattern is Cn, and the similarity is Snx. Equations (1) and (2)
If the formula is satisfied, it is determined that the input pattern belongs to category Cm. If equation (3) is satisfied, the categories Cm and Cn are output and sent to the branch length comparison circuit.
In the branch length comparison circuit 5, branches to be compared are determined using Cm and Cn, and the category to which the input pattern belongs is determined by comparison. Since the branches to be examined differ depending on the category pair Cm and Cn, remember them in advance in a table. Or when converting to hardware, CmC
It is also possible to use n as a parameter and input it to different circuits. For the sake of simplicity, an example will be described in which branches to be examined are stored in the form of a table based on category pairs.

The table shown in FIG. 3 is a parameter table containing category pairs, branch numbers 4, thresholds, etc. In Figure 3, the columns with patterns of "7" and 6T are columns indicating category pairs. In the figure, the columns marked 1 and 2 indicate branch numbers. The next column is the branch to be compared. The number is shown.The rightmost column shows the threshold value.If it is rejected by the judgment circuit and inputted to the branch comparison circuit together with the category pair "7T", the lengths of the second branch and the third branch are compared. .

Letting the length of each branch be L2 and L3, L3-L2〉
When T2, it outputs "7-" When L3-L2≦T1, it outputs "T".

However, Tl and T2 are threshold values given in the parameter table. The above processing will be explained in more detail with reference to the branch length comparison circuit shown in FIG.

It is assumed that the branch table has already been created by the direction encoding circuit, and that the parameter table has category pairs, branch numbers, threshold values, etc. given in advance.

Reference numeral 12 in FIG. 4 is a storage device in which a parameter table is stored, and 15 is a storage device in which a branch table (for simplicity, it is assumed that only branch lengths are stored). The output category pair CmCn from the determination circuit is set in the register 17. The address register 10 is set to the beginning of the parameter table, and the contents are read out and stored in the memory register 13. 13, Cl and C2 are changed to C
A comparison is made using mCn and comparator 18.

If they are not equal, run 11 adders,
The contents of address register 10 are counted up by one. If they are equal, the contents of 11 are transferred to address register 14.
and reads the contents of the branch table into 16 memory registers. The contents of register 16 are transferred to register 19 using signal 103. Next, the contents of 12 are transferred to the address register 14 by the signal 102, the contents of the branch table are read to the register 16, and the gate is opened by the signal 104, and the contents of the register 20 are read.
Move to. The subtracter 21 calculates Li and Lj and transfers the result to the comparator 23. The comparator compares the Li-LJ result with T2. If Li,LJ>T2, the contents of Cm are transferred to the register 22, and if Li-Lj≦T2, the contents of Cm are transferred to the comparator 24.
Compare L1-Lj and T1. When Li-Lj≦T1, the contents of Cm are transferred to the register 22. Li-Lj>
At T1, a reject signal is output. In the above example, the difference in length between the two branches is compared with the threshold value, but it is not necessary to only look at the difference in length between the two branches in order to compare the lengths of the branches.

The length of a single branch may be used for comparison, or the relationship between the lengths of three or more branches may be weighted and used. Alternatively, the length ratio between the two branches of interest in the standard pattern and the length ratio between the two branches in the unknown pattern may be compared, and this may be used as the criterion. In any case, by using the present invention, even patterns that were conventionally rejected as confusing patterns can be correctly recognized by comparing the lengths of the branches.

Note that since the branches to be compared are limited using category pairs, there is no need for the redundant process of comparing the lengths of all branches, which makes it possible to simplify the hardware and improve recognition accuracy.

[Brief explanation of the drawings] Figure 1 is a diagram showing an example of a confusing pattern pair, Figure 2 is a diagram showing an example of a confusing pattern pair;
3 is a diagram showing the contents of a parameter table, and FIG. 4 is a diagram showing a branch length comparison circuit. 12 is a storage device in which a parameter table is stored, 10 is an address register, and 13 is a memory register. 15 is a storage device in which the branch table is stored, 14 is its address register, and 16 is a memory register. 17 is a register for storing the category pair output from the determination circuit; 19, 20, 22 are registers; 18, 23, 24
is a comparator, and 21 is a subtracter.

Claims (1)

[Claims] 1 Calculate the similarity between the input pattern and the standard pattern,
A first determination step of determining which standard pattern the input pattern closely matches, and when a plurality of the standard patterns that closely match are obtained through the first determination step, the input pattern matches the standard pattern. In a pattern recognition method having a second determination step of re-determining which of the patterns it belongs to, the second determination step uses the length of a branch in a predetermined portion of the standard pattern. A pattern recognition method comprising the step of comparing the input pattern with each of the standard patterns and making the re-judgment.