JPH0145100B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a character recognition device that performs character recognition in a hierarchical manner using extracted features.

Conventionally, this type of device is configured to extract all the features that may be used in the decision means in advance in the feature extraction stage from the character pattern, and select and use the features from there in each identification stage. This had the disadvantage that many unused features were extracted, wasting time. Alternatively, the system was configured to extract features directly from the character pattern at each identification stage, so in order to recognize one character, the time-consuming two-dimensional pattern processing had to be performed many times, and a lot of time was spent. It had the disadvantage of being necessary.

In order to solve these drawbacks, this invention first extracts only the basic features from the character pattern, and then uses these basic features to create the necessary features at each decision stage. ,
The drawings will be explained in detail below.

FIG. 1 is a diagram showing an embodiment of the present invention. The character pattern to be recognized written on the form is photoelectrically converted and stored in the storage means 1. The stored contents of the storage means 1 are sent to the feature extraction means 2 through the data line C1. Features are extracted by the feature extraction means 2 and stored in the feature storage means 3 through the data line C2.

The determination in the determination means 4 is made by the first layer identification means 6.
The first layer identification means 6 only selects a candidate category corresponding to the character pattern stored in the storage means 1. The layer identification means 8 selects one category from the candidate categories and outputs it. That is, in the first layer conversion feature creation means 5 corresponding to the first layer identification means 6, the feature storage means 3
The first layer of transformation features are created using the features sent through the data line C3. The first layer conversion features are sent to the first layer identification means 6 through the data line C5, and a plurality of candidate categories are selected as identification results. The candidate category is sent through the data line C6, and the second layer identification means 8 and the second layer conversion feature creation means 7 corresponding to the candidate category operate. In the second layer conversion feature creation means 7, a second layer conversion feature is created using the features sent through the data line C4. The conversion characteristics of the second layer are sent to the second layer identification means 8 through the data C7, and one category is selected in the second layer identification means 8 and outputted through the data line C8.

FIGS. 2 and 3 are diagrams illustrating a specific embodiment of the present invention, and an example in which the invention is applied to number recognition. Characteristics of the character pattern 9 stored in the storage device 1 include the number of loops, the position of the leftmost point of the character pattern 9 in each line examined in the row direction a, and the character pattern 9 in the next line examined in the column direction b. Left difference table 10, which examines the difference between the position of the leftmost point of A right difference table 11 that examines the difference with the position of the rightmost point is extracted by the feature extraction means 2 and stored in the feature storage means 3. In the first layer conversion feature creation means 5, the presence or absence of concave and convex portions on the right side of the character pattern 9 is checked from the right difference table 11 as the first layer conversion feature, and the presence or absence of the concave and convex portions, that is, the conversion feature. Using the loop number, the first layer identification means 6 selects '7' and '1' as candidate categories from the numerical categories 12. A second layer identification means in which the sum of positive differences is obtained from the left-hand table 10 as a second layer conversion feature in the second layer conversion feature creation means 7 corresponding to candidate categories '1' and '7'. 8, the identification result '7' is obtained. At this time, there is no need to obtain the conversion features used by the other second layer identification means 13 and 14.

With this structure, there is no need to extract features that are not necessary for recognition, which has the effect of eliminating wasted recognition time.

Although the above description has been made of the case where the determining means has a two-layer hierarchical structure, the present invention is not limited to this, and may be applied to a case where the determining means has a three-layer or more hierarchical structure.

Furthermore, the features extracted by the feature extracting means are not limited to the features used in this explanation.

As described above, according to the present invention, feature extraction from a character pattern is performed only once at the beginning, and the identification means of each layer uses the contents of the features previously extracted by the conversion feature creation means. By performing calculations or the like to newly create and use necessary features, it is possible to shorten the time required to extract features used in the decision stage.

In addition, since a conversion feature creation means is provided corresponding to each identification means in each layer, for example, in the case of characters 2 and 7, when identifying standard patterns "2" and "7", and when identifying standard patterns "2" and "7", ”, the effective features are different, i.e. characters 2 and 7.
Even if the features used differ depending on the identification means, by providing a conversion feature creation means for each identification means, features can be extracted flexibly and high reading accuracy can be obtained. play.

[Brief explanation of drawings]

FIG. 1 is a diagram showing an embodiment of the present invention, and FIGS. 2 and 3 are diagrams for explaining an example in which the invention is applied to number recognition, which is a specific embodiment.
In the figure, 1 is a storage means, 2 is a feature extraction means, 3 is a feature storage means, 4 is a determination means, 5 is a first layer conversion feature creation means, 6 is a first layer identification means, and 7 is a second layer 8, 13, 14 are second layer identification means, 9 is a character pattern, 10 is a left difference table,
11 is the right difference table, 12 is the numerical category, C1~
C8 is a data line. Note that the same reference numerals in the figures indicate the same or corresponding parts, respectively.

Claims (1)

[Claims] 1. In a character recognition device that extracts features of a character pattern and recognizes characters using the above features by a determining means having a hierarchical structure, a storage means for storing character patterns to be recognized, and a storage means for storing character patterns to be recognized; a feature extraction means for extracting features from a character pattern, a feature storage means for storing the features extracted by the feature extraction means, a determination means having a hierarchical structure including a plurality of identification means; Corresponding to each identification means, a conversion feature creation means for newly creating a feature to be used in the identification means based on the features stored in the feature storage means, and the determination means sequentially converts the features from the upper layer to In the process of narrowing down candidate characters for recognition results by operating the conversion feature creation means and identification means of each layer toward the lower layers, and finally determining the recognition result, the above-described identification is performed at each layer. A character recognition device characterized in that the features used by the identification means are created and used by the conversion feature creation means that corresponds to the identification means.