JPS6326915B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a character recognition method, and in particular, in a handwritten character recognition method, it is possible to identify similar characters among handwritten characters compared with a dictionary, such as by masking the bias. This system performs character recognition by comparing characters, such as 时, with a dictionary.

Conventionally, the method for recognizing patterns such as characters is to read the characters written within the frame of a sheet using an optical reader, apply the photoelectrically converted signal to a frame detection circuit to detect the frame, and then send it to a cutting circuit. Therefore, it is converted into a binary signal of "1" and "0" corresponding to the character background (white part), and the feature of the handwritten character is extracted by a feature extraction circuit. On the other hand, the dictionary is divided into categories according to features and is converted into a binary signal from which features are extracted, and the above-mentioned handwriting is converted into a binary signal and features are extracted.
A known method is to select several types of characters whose compared values are similar, and to recognize, for example, the character with the smallest comparison value as a character written in a frame.

The comparison operation as described above will be referred to as matching in the following description.

That is, in the conventional configuration, such matching is performed by cutting out the handwritten characters 3 in the frame 1 with uniform dimensions as shown in FIG. They were matching.

For this reason, in the case of handwritten characters such as kanji consisting of a bias and a 旁 as shown in Figure 1, there are many characters in the dictionary that have at least a common bias, and the difference in comparison output between these characters is small. , which had the disadvantage of poor recognition accuracy.

The present invention attempts to provide a character recognition method that eliminates the above-mentioned drawbacks, and its feature is to select dictionaries of categories that have similar biases, and to apply the same bias to the dictionaries of the selected candidate categories. On the other hand, the purpose is to mask the biased parts and perform character recognition to match only the blank parts, thereby improving the recognition rate compared to a group of characters with similar biases. An embodiment of the present invention will be described in detail below with reference to FIGS. 2 to 4.

FIG. 2 shows a system diagram of the present invention, in which a document 4 in which handwritten characters 3 are written in a frame 1 is imaged by a vidicon 5, the characters 3 in the frame 1 are read out, photoelectrically converted, and an extraction circuit 6 The extracted signal is extracted and added to the normalization circuit 7 at the next stage. Since the size of characters varies depending on individual habits, the normalization circuit reduces excessively large characters, narrows the line width, and removes noise. The normalized pattern is classified into a plurality of features and extracted by the feature extraction circuit 8.

The feature string extracted and binarized in this way is sequentially compared with the binarized characters stored in the dictionary 10 in advance, that is, matching 9 is performed. Thus, characters similar to the handwritten characters in frame 1 are recognized 11 from the dictionary. In this case, the plurality of candidate characters selected from the dictionary 10 by the above recognition are masked by the mask processing circuit 12, and the second recognition 13 is performed.

That is, the present invention performs matching on the two-dimensional pattern from the videocon in the entire area of the two-dimensionally generated feature sequence, selects a dictionary of similar categories, and matches the input two-dimensional pattern within the dictionary of the selected category. It is a method that matches the left half or right half of a pattern within a frame, and if matching on either the left or right side is clearly similar on average within the dictionary of similar categories than matching on the entire area, it is considered to be similar. , and if it is dissimilar on average, the mask area is not set.If we consider these mask settings for gold-biased sets such as "bell,""iron," and "copper," we get 3. The difference between the two sets is that the ``dou'', ``lost'', and ``dou'' parts of 旁 are different, and the difference in the way of writing the kinbia part is small, even if there is a difference. Then put a mask on the uneven part,
Make sure to match only the straw parts.

Below, the masking method is shown in Figures 3 and 4.
This will be explained with reference to the figures. As shown in Figure 3, the left half of the three patterns made of gold is "0" of the mask 14.
If a mask is selected for this part in the area shown by , no matching is performed and only the area shown by "1" of the mask 14 is matched. That is, FIG. 4 shows an embodiment of the present invention of the mask processing circuit 12 portion, in which a handwritten character 3 for "bell" is drawn in a frame 1 of a document 4, and further includes a recognized gold mark 11. If similar characters ``tetsu'', ``bronze'', ``bell'', etc. are selected in the order of 1st, 2nd, and 3rd place, the 1st place ``tetsu'' and frame 1 are selected.
Give masks 14, 14 to the handwritten character "bell" inside. In this case, when the mask recognizes that the difference in the gold-biased part (left half) is small as described above (distance in equation 1 described later), the gold-biased part is represented by a binary class of "0". A comparison is made between the ``child'' part and the ``lost'' part, that is, the matching area 15, 1.
6, only matching is performed. This matching is performed for the "child" and "lost" parts, that is, if the input and dictionary values are A and B, the absolute value calculation circuit 17 calculates |A-B|, and the absolute value addition circuit 18 calculates the distance α. Calculate (AB) so that it satisfies the following formula (1) α (AB) = _n 〓 ^k=1 |A-B| Do this for k characters. However, n is the number of extracted features expressed in a predetermined number of bits. That is, next, the 2nd place bronze ``dou'' is matched with the input ``dou'', and then the 3rd place ``bell''``dou'' is matched with the input 时 ``dou'', and all of these are matched. The distance α is determined by the minimum value calculation circuit 19, and the minimum value is output to the output terminal 20 as input recognition.

In the above explanation, when the bias distance α is small between the input and the candidate pattern obtained in the first recognition step 11, the bias is masked and only the half is matched, but when the bias between the input and the candidate pattern is not similar In other words, it is clear that when the distance α is large, the time is also matched. When the distance between the input and the candidate pattern is small, the right half may be masked and only the left half may be matched. Since the present invention performs pattern recognition using the method described above, it is characterized in that recognition accuracy can be further improved and recognition rate can be improved.

[Brief explanation of the drawing]

Fig. 1 is a pattern plan view for explaining the conventional pattern cutting method, Fig. 2 is a system diagram showing the method for recognizing characters such as patterns of the present invention, and Fig. 3 is a mask showing the character recognition method of the present invention. FIG. 4, which is a plan view of the explanatory pattern, is a systematic explanatory diagram of the mask processing circuit portion of the present invention. 1...Frame, 2...Cut out, 3...Character, 4...
...document, 5...videocon, 6...cutout circuit, 7...
... Normalization circuit, 9 ... Feature extraction circuit, 9 ... Matching circuit, 10 ... Dictionary, 11, 13 ... Recognition area, 12 ... Mask processing circuit, 14 ... Mask, 17 ... Absolute value calculation circuit , 18... Absolute value addition circuit, 19... Minimum value calculation circuit.

Claims (1)

[Claims] 1. A plurality of features of the handwritten characters 3 in the frame 1 read by an optical reading device are compared with a plurality of features of the characters stored in a dictionary 10 arranged in categories according to features. In a character recognition method for recognizing characters, in the first recognition 11, the handwritten characters 3 within the frame 1 are
The character features in all areas are matched with the character features in all areas stored in the dictionary 10, and based on the results, similar character parts and dissimilar character parts are recognized. A mask is provided for the character part, and only the dissimilar character part is recognized by the dictionary 10 in the second recognition 13.
A character recognition method characterized in that the handwritten character 3 in the frame 1 is recognized by comparing it with a character part stored in the frame 1.