JPH0351981A - character recognition device - Google Patents

Abstract

PURPOSE:To recognize accurate characters speedily by defining only small letters as candidate characters for an extracted character after a small letter is recognized once halfway in a word, and setting the candidate characters while judging character spaces at all times. CONSTITUTION:The character recognition device is equipped with a character input device 102, a character display means 103, a ROM 104 stored with character data for recognition and a word dictionary, and a RAM 105 for storing character images and a position 202 which is a half as high as a small letter is scanned in a word area to extract connection components in eight directions every time a character is met. Features of the extracted character are extracted and compared with all characters in the character data dictionary in the ROM 104 and the most matching character is regarded as a recognition result. When a small letter is judged, candidate characters of following extracted characters are narrowed down to the small letters and when there is a character space more than necessary in the same word, it is considered that the candidate characters of the extracted character may be either of a large letter or small letter. Consequently, characters can be recognized speedily.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention extracts a character area from a document image by inputting a character image of a Western document, etc. using an optical method.
This relates to a character recognition device that converts into code numbers. [Prior art] In recent years, rapid advances in character recognition devices have made it possible to automatically extract character areas from various character images, cut out and recognize individual characters, and automatically create document files. Various methods have been devised to recognize characters. For example, one way to recognize characters is to compare the mesh features and peripheral features of the characters. These methods are described in Research and Practical Application Report No. 34 @ No. 1 pp.
Published in 47-58. The mesh feature expresses the rough shape distribution of the entire character, and the method of extracting the feature is to divide the circumscribed rectangle of the character to obtain nxn small regions. The area of the character portion included in each of the small areas is counted and used as a mesh feature. This mesh feature has nxn data for one character, and the character can be estimated by comparing the area ratio of the character part in a certain area of nxn with data owned as a dictionary. I do. Furthermore, the verifical feature focuses on peripheral information of a character, and the method for extracting the feature is first to find a circumscribed rectangle of a character pattern, and then divide each side of the circumscribed rectangle into n parts. Next, it scans toward the characters from each divided side, and counts the area until it encounters the first character, and then the area until it encounters the next character. By performing the same processing on each divided side, verifical features with nx4x2 data can be obtained, and characters can be estimated by comparing the nx4x2 data with the data held as a dictionary. becomes. The above method is
The features of the extracted characters are extracted, and the features are compared with all the data in the dictionary data for recognition, and the closest match is output as the recognition result. Furthermore, before performing the above recognition, each word is extracted from the extracted character line, and each character is further extracted, but this extraction generally involves counting the marginal distribution in the direction perpendicular to the line direction of the extracted character line. , character spacing and word spacing are estimated from the values of the marginal distribution, and each character is extracted. [Problems to be Solved by the Invention] However, when characters are extracted using the above-mentioned character extraction method, it is not possible to accurately extract characters from a character image as shown in FIG. 5 501. In the character image 501, the characters do not touch each other at all. However, in the direction perpendicular to the row direction, the Y marginal distribution and the O marginal distribution contact each other to form one marginal distribution 502. Therefore, when extracting characters by estimating the character area from the shape of the marginal distribution, either Y and O should be extracted all at once, or they should be forcibly extracted at either position 503 or 504 in Figure 5. I will do it. Therefore, there is a high possibility that an incorrect result will be produced when the extracted character image is recognized.

Also, when recognizing the extracted characters by comparing them with the character data in the dictionary, the recognition process always compares all the character data and outputs the most matching character. The time required becomes extremely long, and as a result, the amount of processing per hour of the character recognition device is reduced, resulting in a drop in performance. Therefore, the present invention is intended to solve the above-mentioned problems, and its purpose is to provide a character recognition device for European and American documents, etc.
The object of this invention is to provide a character recognition device that can accurately and quickly recognize characters using a simple algorithm even in character images with overlapping peripheral distributions in the direction perpendicular to the character line direction. [Means for Solving the Problems] The character recognition device of the present invention includes (1) an optical image input means for inputting a document image by photoelectrically converting light reflected from the paper surface of a European and American document, and character line recognition from the input image. and a means for detecting the position of words and extracting each word, and a character recognition means for extracting each character from the extracted word and converting it into a character code while comparing it with a pre-existing character data dictionary. It is equipped with a means for determining whether a recognized character is an uppercase or lowercase character, and word dictionary data, and when extracting each character, it checks the interval between the extracted character and the character extracted before the extracted character, and determines whether the extracted character is the same. If there is an unnecessarily large spacing between characters within a word, the candidate characters for the extracted characters are both uppercase and lowercase. (2) After the recognized character is determined to be a lowercase letter by the means for determining whether it is an uppercase letter or a lowercase letter, candidate characters for extraction characters are limited to lowercase letters. (3) To extract the above-mentioned characters, scan the vicinity of the center of the height of the character line from left to right or right to left, and extract the connected components of black pixels in 8 directions at the point when a character is encountered. It is characterized by <4> Word dictionary data before outputting the recognized characters! It is characterized by comparing the words in the text with the recognition results and correcting the data. (5) The comparison with the word dictionary data is performed by ignoring, for example, first-class symbols in the word dictionary data and recognized words, and then the words in the word dictionary data are output. [Example] The present invention will be explained in detail below based on an example. Figure 1 shows a block diagram of the character recognition device of the present invention. The character recognition device is a CP that executes processing according to a program.
U101, image input device F for inputting character images into a storage device;
? 1 0 2. Character display means 103 for displaying character recognition results, R containing character data word dictionary for recognition.
OM104, RAM which is a storage device that stores character images
It consists of 105. Hereinafter, the character recognition method of the character recognition device of the present invention will be explained in detail based on the flowchart shown in FIG. 4 and with reference to FIGS. 2 and 3. The character recognition device of the present invention first inputs characters written on paper or the like by an optical method to the RAM 105, which is a storage device, as image data in the image input device W102. Next, extract word regions from the input character image. To extract a word region, first count the peripheral distribution of the input character image in the character line direction. The value of the marginal distribution (not shown) becomes large at the position where a character line exists, and the value of the marginal distribution is small between character lines. Therefore, it is possible to easily estimate the position of a character line based on the value of the marginal distribution. Once the position of the character line is estimated, the marginal distribution in the direction perpendicular to the line direction of the extracted character line is counted. Areas where the value of the marginal distribution (not shown) is large are areas where characters exist, and areas where the value is small are areas where no characters exist. Therefore, by examining regions where no characters exist, the size of word spacing and character spacing can be estimated, and word regions can be extracted. Once a word region is extracted, the next step is to perform recognition while extracting each character within the extracted word. FIG. 2 201 shows the extracted words. In the word, all letters do not touch any adjacent letters. Therefore, it is possible to extract characters by estimating the region 204 where characters exist and the region 203 where no characters exist based on the value of the marginal distribution in the direction perpendicular to the character line, as shown in FIG. 2 205. , there is a possibility that a character image like the one shown in FIG. Then, scan to the right and extract connected components in 8 directions every time you encounter a character. Next, extract the characteristics of the extracted character and store it in the ROM.
104 in the character data dictionary, and the most matching character in the character data dictionary is taken as the recognition result. Also, by recognizing the extracted characters here, it becomes possible to determine whether the extracted characters are uppercase or lowercase.
Normally, in Western documents, if a word changes to lowercase in the middle, subsequent characters are lowercase. Therefore, if it is determined that the extracted character is a lowercase character, it becomes possible to narrow down the candidate characters for subsequent extraction characters to lowercase characters, which makes it possible to speed up character recognition. In the case of 201 in Figure 2, the first character is a capital letter A, so all uppercase and lowercase letters are eligible for the second character. However, by recognizing the second character, it can be determined that the second character is a lowercase character, so that only lowercase characters become candidate characters for subsequent characters, and the speed of character recognition becomes faster. Also, in European and American documents, Figure 2
There is a character image like the one shown in 02. The character image is
In the marginal distribution perpendicular to the row direction, it is interpreted as one word as shown in Figure 2 203. However, there is a - (bar) in the middle, and the next letter is capitalized.
This often happens, and if you try to recognize it as a lowercase letter, you will get it wrong. To do this, it is sufficient to recognize the - in the middle, but in order to perform the recognition at high speed, the present invention
Only one line at 1/2 the height of the lowercase letter is scanned.
Therefore, in a case like the character image 201, there is a possibility that one will not be picked up as shown in FIG. In such cases, in order to avoid mistakes in recognition, the positions of the right and left edges of the character are determined at the same time as character extraction. By comparing this position with the position of the right end of the previous extracted character, the distance between the previous extracted character and the previous extracted character can be determined.
If the interval is extremely large compared to the normal character interval, it can be estimated that something like 1 exists between them, and as a result, it can be determined that the extracted character may be an uppercase letter. When the extracted words are recognized in this way, a word 206 with one removed can be obtained as shown in FIG. 2 206.
Before outputting the word 301 recognized in this way, ROM1
Compare with the word dictionary in 04. The reason for comparing with the word dictionary is to supplement the symbols (not letters) that existed in areas that were extremely spaced apart compared to normal during the above recognition, and the comparison with the word dictionary is to supplement the symbols (not letters) that were present in the areas that were extremely spaced apart compared to normal. Do this by ignoring the symbols. By comparing in this way, it is possible to search for a word 303 in the word dictionary 302 from the recognition result 301, and by outputting the word 303, it is possible to return an accurate recognition result. As mentioned above, in character recognition of European and American documents, etc., we take advantage of the characteristic of Western documents that if a single lowercase letter appears in the middle of a word, the following characters are lowercase letters, and after recognizing the lowercase letters, the extraction character is Since the candidate characters are only lowercase letters, it is possible to significantly shorten the time required for recognition.
In addition, character extraction extracts connected components in eight directions, and at the same time constantly judges the interval between the extracted character and the previous extracted character, so there is no possibility of mistakenly determining candidate characters, and This makes it possible to accurately recognize characters even in character images that overlap in the peripheral distribution in different directions. [Effects of the Invention] As described above, in the present invention, after determining that a recognized character is a lowercase letter during word recognition, recognition is performed by narrowing down candidate characters for extraction to lowercase letters, resulting in faster recognition. ．． Furthermore, even when a word changes to uppercase in the middle of a word, candidate characters are always determined while determining the character spacing, eliminating errors such as recognizing uppercase letters as lowercase letters. Furthermore, since characters are extracted by connecting components in eight directions, characters can be extracted accurately even in character images that overlap in the peripheral distribution in the direction perpendicular to the line direction, and as a result, characters can be recognized correctly. As described above, according to the present invention, high-speed and accurate character recognition is possible. As a result, it has the effect of greatly improving the reliability of character recognition devices that use this method as a structural element.

[Brief explanation of drawings]

Figure 1 is a block diagram of the character recognition device of the present invention. Figure 2,
Figure 3 is a diagram showing how the present invention performs character recognition. 201... Input image 202... Character extraction scanning line 206... Word recognition result 301... Word recognition result 302... Word dictionary 303... Finest word recognition result Figure 4 shows the result of the present invention. Flowchart of character recognition device. FIG. 5 is a diagram showing how character extraction positions are determined based on the shape of the peripheral distribution. that's all

Claims (5)

[Claims] (1) Optical image input means that inputs a document image by photoelectrically converting light reflected from the paper surface of a Western document, and means that detects the position of character lines and words from the input image and extracts each word. a character recognition means for extracting each character from the extracted word and converting it into a character code while comparing it with a pre-existing character data dictionary; a means for determining whether a recognized character is an uppercase or lowercase character; and a word dictionary. When extracting each character, the interval between the extracted character and the character extracted before the extracted character is checked, and if there is an unnecessarily large character interval within the same word, the extracted character A character recognition device characterized in that candidate characters are both uppercase and lowercase letters. (2) The character recognition device according to claim 1, wherein after the recognized character is determined to be a lowercase character by the means for determining whether the recognized character is an uppercase character or a lowercase character, candidate characters for extraction characters are limited to lowercase characters. (3) To extract the above-mentioned characters, scan the vicinity of the center of the height of the character line from left to right or right to left, and extract the connected components of black pixels in 8 directions at the point when a character is encountered. The character recognition device according to claim 1, characterized in that: (4) The character recognition device according to claim 1, wherein before outputting the recognized character, words in the word dictionary data are compared with the recognition result to correct the data. (5) Comparison with the word dictionary data is performed by ignoring symbols such as - (bar) and ' (apo) in the word dictionary data and recognized words, and then comparing the words in the word dictionary data. The character recognition device according to claim 1, wherein the character recognition device outputs a character recognition device.