JPS63201867A - Document image automatic summarization method - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to an automatic document image summarization method for obtaining materials useful for easily understanding document contents, and particularly relates to a document image storage/retrieval system. The present invention relates to an automatic document image summarization method suitable for automatic indexing.

[Conventional technology]

In the conventional technique, as described in Japanese Patent Application Laid-open No. 138670/1984, a summary document is created by providing a means for specifying a location to extract from a document and cutting out the necessary portions.

[Problem that the invention seeks to solve]

The above-mentioned conventional technology does not take into account the need to read documents one by one and specify precisely the extraction points within them, and there is a problem in that a huge amount of cost is required for summarizing a large amount of documents. there were.

Further, there is also a problem in that documents are composed of various elements and there are various means of summarizing them, which cannot be properly summarized.

The purpose of the present invention is to solve the problems of the prior art and to efficiently construct and retrieve a large-scale document image storage and retrieval system.
The purpose of this invention is to provide automatic document image indexing technology for operational use.

[Means for solving problems]

To achieve the above objective, first, the input document image is automatically divided into regions according to its content structure, and each divided region is subjected to an appropriate summarization process prepared in advance. Summary information is concentrated in one image and output.

[Effect]

By automatically dividing a document image into regions according to its content structure, regions can be associated with summarization means necessary for the regions. As a result, documents made up of various elements can be processed flexibly, and the area to be processed by each summarization means can be limited, so automation and efficiency of processing can be easily realized.

〔Example〕

Hereinafter, one embodiment of the present invention will be described in detail.

FIG. 1 is a block diagram showing the processing flow of an automatic patent publication abstract image creation method according to the present invention.

A gazette document input as an image in the gazette image input section 11 is first extracted in a component extraction section 12 for each character or figure in units of circumscribed rectangular frames.

An example is shown in FIG. Note that this can also be implemented using a known processing algorithm.

As shown in FIG. 3, the element arrangement analysis unit 13 extracts the following feature amounts from the elements 32 cut out in units of the circumscribed rectangular frame 31.

■, X coordinate of the upper left point of the circumscribed rectangle: x1■, X coordinate of the upper left point of the circumscribed rectangle: Y1■, Horizontal size of the circumscribed rectangle: WI■, Vertical size of the circumscribed rectangle :HI■, X-axis projection marginal distribution of element: (pt(m))(m=1...
・, Wt) ■, Element's y-axis projected marginal distribution: (Qi(n)) (n=1
．． ..., Hl) Here, the subscript i indicates the feature amount of element i. Among these, the document shown in FIG. 1 is roughly divided into the following three areas using ■ to ■.

A, bibliographic matter area 21 (first page) B, specification text area 22 (first page~) C0 specification drawing area 23 (~last page) A, B, and C appear in this order, so A, B, and C appear in this order. B.

If B and C can be distinguished, the above area division becomes possible.

The boundary between A and B exists as a large blank area in the first page. Therefore, this is SP+= (Xt Xi+t)2+100(Yt
It is sufficient to detect the element j that maximizes "Yt+t)". Note that weighting is given in the y direction in consideration of the characteristics of the blank area.

On the other hand, the boundary between B and C is W, which indicates the size of the circumscribed rectangular frame.
Alternatively, it can be easily detected at the element i where Hl exceeds the allowable value for the first time. Alternatively, the irregular arrangement of elements, which is a feature of a non-text area, can be expressed as the upper left point of the circumscribing rectangle (x, +, Yl
) can also be detected.

Next, detailed division and summary processing shown in blocks 14 to 16 are applied to the three divided regions, respectively.

The bibliographic item extraction unit 14 extracts bibliographic items from the bibliographic item area 21. Prior to describing an item, a publication displays a specific character string (hereinafter referred to as a headword) that clearly expresses the content. An example is shown in FIG. Therefore, by searching for this headword 41, the location of each bibliographic item 42 can be limited (the area can be divided in detail). After that, only bibliographic items are extracted with high accuracy based on the relative positional relationship with the headword. Note that there may be multiple headwords (for example, N
Since the character string is made up of several characters, the probability of misjudging all characters is extremely small.Using this combination effect, even if the search accuracy for each character is insufficient, it is possible to perform a high-precision search as a whole. becomes. In other words, there is no need to use expensive equipment such as OCR. For example, the registered entry word (at); t=1. A match between ..., N and the candidate character string (bt) in the processed document; t=, t, ..., N is determined as N consecutive occurrences of similar at"bt for each character. In this case, , a simple and fast headword search can be carried out by performing similarity judgment for each character using the marginal distribution using the following formula.In other words, if kuK・((S + + S J)/2) is satisfied. If so, it is determined that characters i and j are similar.Here, C
is the mutual shift amount for correcting the positional shift, and Wl
and W.J., Pl and p.

SL and S are the horizontal size, X-axis projected peripheral distribution, and area of the circumscribed rectangular frames of i and j, respectively. Also, K
is a proportionality constant determined by the entire document.

Note that the same applies to the similarity determination using the y-axis projected marginal distribution, and by performing the determination in two stages, it is possible to improve accuracy without increasing the load too much.

The frequently appearing character string extraction unit 15 extracts frequently appearing character strings as keyword candidates from the specification text area.

If the entire specification is targeted, the processing efficiency will decrease and the possibility of extracting non-keywords will increase. This is limited to Scope J. This is based on the two headwords “Claims” and “Detailed Description of the Invention.”
It is extracted as an area surrounded by. As a method for extracting frequently occurring character strings from this limited area, Japanese Patent Application No. 61-288775 filed by the present applicant entitled "Frequently occurring character string extraction method"
There is.

The main drawing extraction unit 16 extracts main drawings from the specification drawing area. Generally, there is a tendency to use Figure 1 as the main drawing, so of the circumscribed rectangular frames of a certain size or more that are considered to be one figure, the one located at the upper left (earliest) is extracted as the main drawing.

Since the sizes of the main drawings vary depending on the publication, similarity transformation is performed so that the long sides of the circumscribed rectangular frame are smaller than a certain size.

Next, in the abstract element editing section 17, as shown in FIG. 5, bibliographic items 51. Frequently occurring character string 52. An abstract image is created by editing the three abstract elements of the main drawing 53 into one image.

The abstract image output unit 18 displays the created abstract image on a CRT display or stores it in a storage device such as an optical disk.

According to this embodiment, since the structure of a document can be accurately grasped by searching for headwords, more efficient and highly accurate summarization processing can be realized.

Furthermore, in the above embodiments, if code conversion processing using OCR or the like is applied to the divided specification text areas, etc., even more flexible summary processing can be performed. For example, existing software such as Japanese language processing developed on the code can be used to perform semantic analysis and translation.

〔Effect of the invention〕

According to the present invention, by automatically dividing an input document image into regions according to its content structure, it is possible to limit the processing target region and prepare an appropriate summarization process for each region. This has the effect that automation of the image summarization device can be realized efficiently and with high accuracy.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of an embodiment of the method for creating an abstract image of a published patent publication according to the present invention, and FIG. 2 is a diagram showing an example of the components of the publication extracted by the component extracting section of the abstract image creation method. Figure 3 is a diagram showing the feature amounts of the above-mentioned components, and Figure 4
Figure 5 shows the positional relationship between headwords and bibliographic items.
The figure shows an example of an abstract image.

Claims (1)

[Claims] 1. A document is input as an image, the input document image is divided into regions according to its content structure, and a predetermined summarization process is individually prepared for each divided image region. administer,
An automatic document image summarization method characterized in that summary information extracted from each image area by the summarization process is edited into one image and output. 2. Automatic summarization of document images according to claim 1, characterized in that, after converting the image into another representation for the divided image area, a predetermined summarization process is performed on the representation. method. 3. Input a document as an image, register in advance a character string that will become a heading indicating the contents of the document, search for the character string from the input document image by image processing, and add the searched character string to the character string. An automatic document image summarization method characterized in that the document image is divided into regions based on the document image.