JPH04147391A - Text recognition apparatus - Google Patents

Abstract

PURPOSE:To improve a correcting effect of a recognition result in the recognition process for a text image by displaying layout information and a result of character recognition and by reprocessing the text recognition based on layout correcting information. CONSTITUTION:Layout analysis section 2 extracts a layout of text image. A character recognition section 3, based on layout information, recognizes a character section, and outputs a result of recognition in conjunction with object characters. A knowledge processing section 4 carries out context processing from a result of character recognition. The correcting interface section 51 of a correcting section 5 displays layout information and a result of character recognition and makes layout correction possible. A reprocessing instruction section 52, based on information from a correcting interface section 51, instructs the knowledge processing section 4 for reprocessing. The operator uses an instruction device such as keyboard/mouse 511 to carry out the text correction while watching a recognition result display section 512.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to an apparatus for recognizing document images and creating electronic documents.

[Conventional technology]

With the development of character recognition technology in recent years, various OCR (Opt)
ical Character Reader) has been developed.

Until now, methods that use the run length of both black and white pixels have been used to extract elements such as characters, diagrams, or photographs that make up a page without using prior information about the format.
one that uses pixel density within a window of constant size;
Some methods use the proximity linear density method. The main purposes of these methods are to separate grayscale figures and binary figures, and to separate graphic areas and character areas, on the premise of image storage, editing, or image transmission.

On the other hand, a method based on tracking connected components of black pixels has been reported as a method for extracting character strings based on character recognition. There is also a method in which rough information about the format, such as column positions, is given in advance and page components are extracted using this information as clues.

Furthermore, Honda, Bosou, "Paper component extraction method that does not depend on format specification information," Proceedings of the Institute of Electronics and Communication Engineers, 83/1. Vo
l, J66-DNo, 1. In papers describing pH1 to pH8, field separators, i.e., prior information about the position on the paper, are used to collect information from the paper, which consists of a linear figure that forcibly divides the area on the paper, a heading, and a text string. A method has been proposed to automatically extract these page components without using any information.

In this way, when automatically recognizing the arrangement of page components, that is, the layout, there are cases where the recognition is erroneous. For this reason, humans had to make corrections to the automatically recognized layout.

In conventional document recognition devices, when correcting the recognition results, the recognition results are corrected by correcting only the layout when correcting the layout, and correcting only the text when correcting the text part. . Therefore, even if a character recognition error occurs due to an incorrect recognition of the layout structure, all you need to do is correct the layout.
, the text part was not automatically corrected.

To explain with a specific example, for an original image consisting of two columns of text as shown in Figure 3(a) and characters such as line numbers in between, the The layout analysis should be done by dividing into two columns and the characters in the middle, but the characters in the middle and the character lines in the blocks at both ends are horizontally close to each other, as shown in Figure (C). In some cases, the layout is analyzed such that the intermediate character part is merged with the line of the adjacent column. In this case, as shown by the area 6 surrounded by a broken line in Figure (d), one line in the left column, the intermediate character part, and one line in the right column are considered to be one continuous line. Because context processing is performed, it is originally [Qi]. The part where context processing should be performed, such as “Tomorrow 10 Loaf,” is shown in the figure (
As shown in e), context processing is mistakenly performed as ``Go J'' in 1007 tomorrow.

Note that the context processing is performed in this way because, as will be described later, recognition of numbers may be prioritized.

In this way, if layout analysis is performed incorrectly, knowledge processing, which involves performing context processing based on character recognition results to create sentences that make sense, will not work properly, as shown in Figure 3 (e). ) will output incorrect recognition results as shown in the figure below.

In such cases, the operator would have to manually modify the layout, but in the past, even if accurate layout analysis had enabled knowledge processing of the character recognition results to yield a correct answer, the operator would have to manually modify the layout. Even if I corrected it, the character recognition results were not automatically corrected.

[Problem to be solved by the invention]

As described above, in the conventional document recognition apparatus, even if a mistake in layout analysis is corrected by the operator, it does not affect the character recognition result in any way and does not function properly.

The present invention has been made to solve the above-mentioned problems, and an object of the present invention is to improve the efficiency of correcting recognition results.

[Means to solve the problem]

In order to achieve the above object, the document recognition device of the present invention includes: an image input section that inputs a document as an image; a layout analysis section that extracts the layout of the document image input by the image input section; Character recognition is performed based on the layout information obtained from the layout analysis section.
7i! a character recognition unit that outputs candidate characters along with recognition results; a knowledge processing unit that performs context processing based on the character recognition results of the character recognition unit; and a knowledge processing unit that performs context processing based on the character recognition results of the character recognition unit; The present invention is comprised of a modification interface section that displays the results and allows modification of the layout, and a reprocessing instruction section that instructs the knowledge processing section to perform reprocessing based on the layout modification information from the modification interface section. Features.

[Effect]

According to the present invention, a general document is input as a digital image from the image input section, and the layout analysis section extracts the layout of the original image input from the image input section. On the other hand, the character recognition unit performs character recognition processing that also outputs candidate characters.

Next, the knowledge processing unit performs context processing on the character recognition results including the candidate characters to create a meaningful sentence.

Next, the correction interface section displays the results of the character recognition section that have passed through the layout analysis section and the knowledge processing section on the recognition result display section, and the operator makes corrections using an instruction device such as a keyboard/mouse. When the layout information is modified as a result of this modification, the reprocessing instructing section feeds back to the knowledge processing section based on the layout modification information, and displays the new knowledge processed result on the recognition result display section. In this way, the target document is created by repeating the correction process.

〔Example〕

Hereinafter, features of the present invention will be specifically explained based on embodiments shown in the drawings.

Fig. 1 is a configuration diagram of the document recognition device in this embodiment, and Fig. 2
The figure is a configuration diagram of a correction section used in the document recognition device.

The document recognition device includes an image input unit 1 that inputs a document as an image, a layout analysis unit 2 that extracts the layout of the input document image, and a recognition result that recognizes characters based on the obtained layout information. A character recognition unit 3 that also outputs candidate characters, a knowledge processing unit 4 that performs grammatical processing etc. from the result of character recognition, and a modification interface unit 51 that displays layout information and the results of character recognition and allows modification of the layout. and a reprocessing instruction section 52 that instructs the knowledge processing section 4 to perform reprocessing based on this correction information.

Further, the modification interface section 51 and the reprocessing instruction section 52 constitute a modification section, and the modification interface section 51 displays the recognition result of the keyboard/mouse 511 which is an instruction device and the character recognition section 3. It consists of a recognition result display section 512 and a correction processing control section 513 that receives and receives the results of character recognition. Note that the correction processing control unit 5
13 has a function of creating a document based on input from the keyboard/mouse 511, that is, a word processor function.

Next, the operation of the above-mentioned document recognition device will be explained.

First, a document image inputted by the image input section 1 and binarized is analyzed by the layout analysis section 2. The analysis method is based on the above-mentioned reference "Paper component extraction method that does not rely on format specification information"
The analysis method proposed in et al., i.e., field separator (a linear figure that forcibly divides the area on the paper)
Using a method that automatically extracts these page components from a page consisting of text, headings, and text strings without using any prior information regarding their position on the page, we can extract text-only areas and The character recognition unit 3
Character recognition is performed using

A known method is used for character recognition. In addition to the first recognized character, the character recognition unit 3 also outputs a plurality of candidate characters that are likely to be the same character. The knowledge processing section 4 performs grammatical processing and the like on the recognition result output from the character recognition section 3, and performs a process of converting the character string into a character string that conforms to the prepared grammar.

An example of this process will be described using the image shown in FIG. 3 as an example. Assume now that the recognition results including candidate characters output from the character recognition unit 3 for the original character string are as shown in Table 1.

Table 1 As shown in Table 1, from the recognition results including candidate characters output from the character recognition unit 3, the character immediately before the character of interest among the character strings with the highest probability If is a number, there is a high probability that the character currently being focused on is also a number. Therefore, if the character with the highest probability is not a number but the character with the second highest probability is a number, the second-highest character is output as the recognition result. Do the same thing for kanji, hiragana, and katakana other than numbers. By this process, the character string "10" is displayed in Table 1.
If the character string ``10 ROKU'' is composed of only the characters with the highest probability of recognition processing, it will be correctly recognized as the character string ``10 ROKU''. Here, if rO-7J is the candidate character with the second highest probability for the recognition result "loaf", if the knowledge processing described above is performed, the intermediate character, that is, the line number "1"
0'' to generate the character string rloo-7J. In this way, the results of the character recognition section 3 are transferred to the knowledge processing section 4.
Automatically correct by.

However, this process may occur if the character string to be processed is not correctly extracted, that is, if the processing by the layout analysis unit 2 is not analyzed as the intended layout.
As shown in FIG. 3, this results in a process that lowers the recognition rate. The results of layout analysis, that is, the type of block, the position and size of each character, the results of character recognition using knowledge processing (including candidate characters), and the position of the recognized character in which block it corresponds to. A table showing whether or not there is a fish is also sent to the correction section 5.

In the correction section 5, first, the correction processing control section 513 receives the results of the layout analysis and the character recognition results after knowledge processing in the correction interface section 51, and displays the layout and character recognition results on the recognition result display section 512. I do.

Then, while viewing the recognition result display section 512, the operator makes corrections using an instruction device such as the keyboard/mouse 511.

When displayed, it is possible to make it easier for the operator to understand the type of block by color-coding the blocks, such as text-only blocks in blue and graphic blocks in red.

It is also possible to distinguish between blocks that have been modified and blocks that have not been modified by color. The modification procedure begins with modifying the layout that is easy for the operator to see and understand. Modifications change the type of block (text-only block, figure block, etc.), its position, size, etc. As a result, the table which contains information such as the position and size of blocks, characters included in the block, etc., held by the correction processing control section 51, is rewritten.

Then, when the layout modification is completed and the desired layout has been created, the reprocessing instruction unit 52 instructs the knowledge processing unit 4 to reprocess the changed layout portion.

As a result, if a layout analysis as shown in Figure 3(C) is performed on a document with a layout as shown in Figure 3(a), if the operator makes modifications as shown in Figure 3(a), the knowledge The processing unit can correctly perform knowledge processing even on the area 6 for which the layout has been analyzed incorrectly.

Finally, errors in the character recognition results are corrected using the word processing function of the correction processing control unit 513.

Through these processes, the document desired by the operator is created.

Although the above example describes an embodiment only for document creation, it can also be used for inputting slips.

Furthermore, in the above embodiment, the knowledge processing is instructed to reprocess after all layout corrections are completed, but it is also possible to instruct the knowledge processing to reprocess each time each layout correction is made. is also possible.

Further, in the above embodiment, only reprocessing to knowledge processing has been described, but reprocessing to character recognition is also possible.

By the way, a reprocessing instruction for character recognition is issued when extraction of a character, which is the smallest unit, fails during layout analysis.

To be more specific, if accurate character extraction fails for the character ``林'', the characters will be recognized as ``木'' or ``木''. In this case, a reprocessing instruction is given to the character recognition unit by correcting the two rectangles surrounding the characters ``木'' and ``木'' as one character.

〔Effect of the invention〕

As explained above, according to the present invention, knowledge processing is completely performed since the results of layout correction are processed in the knowledge processing unit, and the efficiency of character correction operations when creating a document is significantly improved. It has the effect of being

[Brief explanation of the drawing]

Fig. 1 is a block diagram of a document recognition device showing an embodiment of the present invention, Fig. 2 is a block diagram of a correction section used in the same document recognition device, and Fig. 3 is an explanatory diagram of problems in the conventional example. be. 2. Image input section 2: Layout analysis section 3: Character recognition section 4: Knowledge processing section 5: Correction section 51: Correction interface 511: Keyboard/mouse 512: Recognition result display section 513: Correction processing control section 52
; Reprocessing instruction unit 6: Incorrect layout analysis area Patent applicant Fuji Xerox Co., Ltd. Agent d) (e)

Claims (1)

[Claims] 1. An image input section that inputs a document as an image; a layout analysis section that extracts the layout of the document image input by this image input section; and a layout analysis section that extracts the text part based on the layout information obtained from this layout analysis section. a character recognition unit that recognizes characters and outputs candidate characters along with the recognition results; a knowledge processing unit that performs context processing based on the character recognition results of the character recognition unit; and a knowledge processing unit that performs context processing based on the character recognition results of the character recognition unit; It is composed of a correction interface section that displays the recognition results and allows the layout to be corrected, and a reprocessing instruction section that instructs the knowledge processing section to perform reprocessing based on the layout correction information from the correction interface section. A document recognition device characterized by: