JPS6349822B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a pattern recognition method, and more particularly to an improvement in a character recognition method using a structural analysis method.

In general, the recognition process involves parametric,
Or it contains discrete choices. For example, selection of the binarization level, selection of whether or not to remove small-area connected components included in the pattern, and selection of which small components to remove if removed (see Japanese Patent Publication No. 17698/1983). ), the choice of whether or not to perform thickening processing as a compensation measure to read even broken patterns, and if so, the number of bits to be thickened.

The character recognition method using the structural analysis method further includes several selections in the structural analysis process. Although there are cases where this option can be selected without ambiguity, ambiguity often remains. If ambiguity remains, it is necessary to prepare dictionary entries (standard patterns) for each option so that any option can be selected.

Usually, the standard (or most reliable) option is selected first, the recognition process is run through, and
If the obtained result has a sufficiently high degree of similarity, that result is used as the answer, and if it does not have a sufficiently high degree of similarity, the least reliable option among the selected ones is changed and the recognition process is retried.

In this way, recognition processing can be seen as having multiple processing paths. The selection of this processing path is determined by the processing parameters at each stage.

These retrials have conventionally been performed from the viewpoints of "hair removal,""noiseremoval,""thickeningprocessing," etc., and are well known. These processes are collectively called "perturbation."

By using these methods, the noisy alphabetic letter "A" as shown in FIG. 1 or the truncated number "9" can be read correctly. However, according to this method, in the case of Figure 2 a and b, for example, the small area component 3 of the katakana ``so'' is regarded as noise and removed, so that the katakana ``no'' is
There is also a problem that the small loop 4 in the number "9" is collapsed and misread as the number "1".

Conventionally, in order to solve this problem, there is a method in which the determination threshold is changed depending on the type of perturbation applied to eliminate misreading. (Special application 1986-36075
According to this method, the reliability of the recognition method can be improved in the sense that it eliminates misreading that may occur when attempting to repair small noises or small cuts. However, with this method, for example, large noises 5 as shown in Fig. 2c and d,
It is difficult to actively remove noise 6 that will not result in misreading even if removed. This is due to the fact that penalties such as threshold changes depending on the type of perturbation are uniformly given to each category.

Therefore, an object of the present invention is to solve the above-mentioned problems, remove noises 5 and 6 in FIG. 2c and d when they can be removed, and remove isolated components 3 in
In cases where it should not be removed, such as in cases where it should not be removed, it is not removed to improve recognition accuracy against noise and character deformation, and thickening processing allows character patterns with large cuts and small irregularities to be recognized. It is an object of the present invention to provide a recognition method that can recognize even the above-mentioned patterns with high reliability while eliminating misreading that may result from thick processing. The present invention will be described below using character patterns as an example, but the present invention is not limited to character patterns and can be applied to recognition of various patterns.

First, the principle and features of the present invention will be explained.

The present invention has been made based on the following three facts (F1 to F3).

(F1) By removing large isolated components and thickening a large number of bits, the range of character patterns that can be recognized with the same dictionary capacity (standard number of patterns) increases considerably.

(F2) Misreadings that occur when implementing the above measures are
A fairly limited number of dictionary entries (standard pattern)
Occurs only in

(F3) There is a fairly strong correlation between the dictionary item in question and the type, degree, and content of perturbation that results in misreading, and it is possible to describe these relationships in advance.

First, regarding (F1), consider the case where we are trying to recognize low-quality character patterns (patterns made with ballpoint pens or thin writing instruments, patterns with cutting errors, free handwritten characters, etc.) using a structural analysis recognition method. , easily understood.

(F2) can be shown experimentally, but it can also be understood from Figure 2. In the structural analysis recognition method, dictionary entries (standard patterns) are prepared for each character transformation.
g and h are dictionary entries (standard patterns) for the number "3". Of these, only g is the standard pattern that may be erroneously recognized by removing the large noise 21 in the unknown character pattern e. Moreover, the area that must be removed is often a very limited area. In this example, only the area around area 22 in g in the figure is an area where removal is prohibited.

In this way, the number of dictionary items that are problematic is generally limited, and the range can be narrowly limited.

(F3) is also demonstrated through preliminary recognition experiments using the created dictionary. By recognizing 500,000 to 1,000,000 characters containing various deformed character patterns and automatically analyzing the causes of misreading, we can identify the dictionary entry number in question, the type of perturbation, the degree and content of the perturbation. It is possible to collect information regarding the correlation between

Therefore, based on these three facts, the present invention controls the recognition process based on information in the processing path on the way to obtaining the recognition result, thereby reducing the penalty caused by perturbation. The process is equivalent to making it variable for each category, and as a result, highly accurate recognition can be achieved with a smaller number of standard patterns than before.

Hereinafter, the present invention will be explained in detail with reference to Examples.

FIG. 3 is a block diagram of one embodiment of a character recognition device using the method of the present invention. This will be explained along the flow of recognition processing.

The unknown character pattern 10 is observed by the observation unit 100, and a binary pattern 11 is obtained by binarization processing. During binarization, a slice level exists as a parameter, and this parameter is denoted as _P1 .

The binary pattern 11 is thickened by the thickening processing section 200. For thicker bits, the number of thicker bits exists as a parameter, and this parameter is written as _P2 . However, the number of thick bits _P2 may be 0. In other words, even if the data cannot be thickened, thickening processing is performed by virtually setting the number of thickening bits to 0.

The binary pattern 12 that has undergone the thickening process is input to the structure analysis section 300, and is subjected to processes such as contour extraction and contour division, as will be described in detail later. The divided contour parts are called segments. Ultimately, the structure analysis unit 300 has the role of dividing the binary pattern of the unknown input pattern into segments and assigning unique numbers to each segment in a predetermined order.

The feature extractor 400 receives the output of the structure analyzer 300 and extracts features that describe the shape of each segment and features that describe the relative positional relationship between the segments.

The pattern matching unit 500 receives the output of the feature extraction unit 400 and evaluates the similarity between the output feature amount and each dictionary item indicating a character concept in the dictionary 900.
Similarity is measured by a distance determined by a predetermined definition formula. In the end, the pattern matching unit 500 inputs the features of the unknown pattern and calculates the distance from each dictionary entry.
N from the smallest distance value (N: an integer greater than or equal to 1)
Then, N pairs of dictionary entry numbers and corresponding distances are output.

The validity testing unit 600 receives the parameters 20 sent from the observation unit 100, the thick processing unit 200, and the structure analysis unit 300, and the pair 1 of dictionary item number and distance.
5, the validity of each dictionary item is tested while referring to the prohibition table 1000. Table of prohibitions 100
The details of 0 will be described later. The pair of dictionary item number and distance whose validity is denied is discarded, and only the pair of dictionary item number and distance that remains becomes the output 16 of the validity test unit 600.

Upon receiving this output 16, the strategy determining unit 700 determines whether to make a final determination or to change the parameters at a certain stage and retry (perturb) the recognition process. When retrying (perturbing), the command 50 is used to transmit to the relevant processing unit which stage and with what parameters the retry is to be performed.

If the strategy determining unit 700 determines that it is OK to make the final determination, among the pairs of dictionary item number and distance,
Only the pair with the minimum distance within each category is selected and transferred to the final determination section 800.

The final determination unit 800 inputs two pairs of category names and distances from the strategy determining unit 700, starting from the one with the smallest distance, and determines that the minimum distance is smaller than the absolute threshold δ, and the difference from the next smallest distance is smaller than the relative threshold ε. Only when the distance is large, the category name that provides the minimum distance is used as the recognition result, and in other cases, the final output 20 is obtained using the code that indicates unrecognizable as the recognition result.

Now, the main processing contents will be explained in detail.

The observation unit 100 and the thickening processing unit 200 can be easily realized using known techniques, so their explanation will be omitted. (For example, ○...Jyoga Gen

Claims (1)

[Claims] 1. Observing an unknown pattern and converting it into an electrical signal,
Perform preprocessing on the observed unknown pattern, evaluate the similarity between the pattern corresponding to the preprocessed unknown pattern and a standard pattern registered in the dictionary, and make a predetermined judgment based on the evaluation result. In a pattern recognition method in which the unknown pattern is recognized through a series of recognition processes, a predetermined rule related to the preprocessing is assigned to each standard pattern in advance, and the unknown pattern is When evaluating the similarity between a pattern corresponding to a standard pattern and a standard pattern, it is determined whether the preprocessing performed on the unknown pattern conforms to the above rules of the standard pattern, and thereby the validity of the evaluation result is verified. Test the gender,
A pattern recognition method characterized in that a predetermined control is performed on the series of recognition processes based on the test results. 2. Paragraph 1, wherein the predetermined rule specifies whether or not at least one standard pattern extracted based on the evaluation result is allowed to be used as a candidate pattern for recognition. pattern recognition method. 3. The pattern recognition method according to item 1, wherein the standard pattern registered in the dictionary includes control information for performing the series of recognition processes. 4. The pattern recognition method according to any one of items 1 to 3, wherein the predetermined control is a control for retrying the series of recognition processes based on the test result.