JP2005208979A - Featured value extracting device and method and document filing device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To solve the problem that when a pixel value is accumulated as it is in a projecting direction, the base color of paper or the background color of an original and image components and ruled line components inserted into a document are also accumulated, and character components in a projection waveform are not clear, and the extraction of any featured value can not be correctly operated. <P>SOLUTION: Projection waveform is generated by a projection waveform generating part 21 from inputted multi-value document data. Concretely, the pixel values of each document data are accumulated in horizontal and vertical directions so that projection waveform data can be generated. The projection waveform data are binarized by a binarizing part 22, and a binary data sequence constituted only of a black region and a white region are acquired. Then, the length of the continuous black regions in the binary data sequence is all calculated by a frequency distribution analyzing part 23, and the value whose appearance frequency is higher than a predetermined threshold value among those values is set as the featured value of the document data. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a feature amount extraction device, a feature amount extraction method, and a document filing device, and more particularly to a feature amount extraction device that extracts feature amounts such as document layout, character size, and line spacing using a projected waveform of a document image. The present invention also relates to a document filing apparatus that determines a delimiter position of a document unit by using the feature quantity extraction method, and manages document data for each document.

  In recent years, when a document composed of one or more paper originals is efficiently digitized over a plurality of copies, a process of continuously reading the originals using a scanner device having an automatic paper feed function has been widely performed. . At this time, in order to manage the read document image data (document data) for each document, it is necessary to detect a break between the documents using some method.

  In order to detect a break between documents, conventionally, a character recognition area portion set in advance is cut out from image data read by a scanner device, and character recognition is performed. Based on the result of character recognition, The document separation is determined (for example, see Patent Document 1), a plurality of documents are read at once by a scanner device having an automatic paper feed function, and the feature amount of the read original image is calculated, and this calculation is performed. The document unit break is determined based on the feature amount (see, for example, Patent Document 2).

  In any of these conventional techniques, document separation is automatically determined without inserting a document separation document (for example, a blank document) between each document in advance or modifying the document to be document separation. Therefore, the burden on the user can be greatly reduced. In particular, with the conventional technique according to Patent Document 2, there is no restriction on the format of the target document, and it is possible to deal with a wider variety of documents. On the other hand, the prior art according to Patent Document 1 can only deal with a document of a specific format.

  By the way, in the prior art according to Patent Document 2, as examples of the feature quantity that is the basis for the delimiter position determination, the hue used for the image data, the number of screen lines, the layout of the document, the direction of character combination, etc. . However, in general office documents, there are many text-only originals that do not include images, and thus feature quantities such as the hue of the image and the number of screen lines cannot often be used to determine the separation position. In addition, since horizontal writing is common, the direction in which the characters are combined is often not useful for determining the document break position.

  On the other hand, feature quantities such as document layout, character size, and line spacing are generally different for each office document. Therefore, it is an effective technique to use these feature amounts for determination of the document break position. As a method for extracting feature quantities such as document layout, character size, and line spacing from a document image, it is common to form a projection waveform in the vertical and horizontal directions of the document image and use the projection waveform. .

  Specifically, the document image is binarized, projected in the vertical and horizontal directions, and a threshold value process is performed on these projection waveforms to determine a character area / blank area. The size and line spacing are detected (for example, see Patent Document 3). Also, binarization of the document image is performed, and then a line is cut out by projecting in the horizontal direction, and then each character is separated by projecting in the vertical direction in the image of each separated line. The character size is detected by correcting the separation of the characters by taking the horizontal projection again (see, for example, Patent Document 4).

  In extracting the character size, first, the document image is binarized, and then the projection is performed in the horizontal direction to cut out the rows, and then the projection is performed in the vertical direction in the separated images of each row. Then, the character size is extracted by taking a horizontal projection again for each character and correcting the character separation (see, for example, Patent Document 5). In addition, information on blocks circumscribing each character or character string is extracted to calculate a standard character block size, and this is compared with the circumscribing block information to continuously detect character blocks. By calculating the projection distribution and its frequency distribution, and calculating the optimum threshold for separating character blocks into circumscribed blocks of one character size from the frequency distribution, the accuracy of character size detection in handwritten characters is improved. (For example, refer to Patent Document 6).

Japanese Patent Laid-Open No. 10-21380 Japanese Patent Laid-Open No. 2002-24258 Japanese Patent Publication No. 7-1111738 JP-A-5-89283 JP-A-5-89283 JP-A-7-98747

  As described above, in each of the related arts according to Patent Documents 3 to 6 using the projection waveform, as shown in FIG. 10, the read document image is first binarized, and the projection waveform for the binarized image is obtained. The feature amount is extracted based on the projected waveform. Therefore, the processing result is greatly affected by the binarization threshold. However, even if the binarization threshold value is set to an appropriate value, the binarization threshold value still depends on the background color of the paper, the background color of the document, the presence / absence of an image inserted in the document, the character color / density. Therefore, it is difficult to obtain a stable processing result.

  As shown in FIG. 11, the above binarization level problem can be solved by taking a multi-value projection without binarization. However, if the background color of the paper or the background color of the document is dark, the image is inserted in the document, the character density is low, or the line is short, etc. The difference in projection is not clear. Further, in the case of a document image containing a lot of low density noise, if binarization processing is performed, the noise is determined to be a white level, and as a result, the influence of noise can be eliminated. However, if projection is performed with multiple values, noise components are also accumulated, which makes processing difficult in this respect.

  That is, in the conventional technology that uses a projection waveform to detect a feature amount such as a character size or line spacing on a character image, if the pixel values are accumulated as they are in the projection direction, the background color of the paper or the background color of the document Since the image components and ruled line components inserted in the document are also accumulated, the character components in the projected waveform are not clear. As described above, when the character component in the projected waveform is not clear and it is unclear which part of the projected waveform corresponds to the character part, the process of detecting the feature amount such as the character size and the line spacing from the projected waveform. Cannot be performed correctly. In addition, the character size extracted from the document is not necessarily a feature amount representing the document.

  The present invention has been made in view of the above problems, and the object of the present invention is to influence the influence of disturbances such as the background color of the paper, the background color of the document, the image component inserted in the document, and the ruled line component. To provide a feature amount extraction device, a feature amount extraction method, and a document filing device using the feature amount extraction device, which can correctly extract feature amounts such as character size and line spacing without receiving them. .

  In order to achieve the above object, in the present invention, projection waveform data is generated from input multi-valued document data, and the generated projection waveform data is binarized. Then, the frequency distribution is analyzed by obtaining all the values indicating the characteristics of at least one of the continuous black area and the white area in the binarization result, and the analysis result is used as the feature amount of the document data. Alternatively, the document unit break is determined based on the feature amount obtained in this way.

  When extracting feature quantities such as character size and line spacing on a character image using the projection waveform, the projection waveform data is generated directly from the multivalued document data, and the generated projection waveform data is binarized. Thus, a binary data series consisting of only black areas and white areas is obtained. Then, the frequency distribution is analyzed by obtaining all the values indicating the characteristics of at least one of the continuous black area and the white area in the binary data series, and the analysis result is used as the feature amount of the document data. An effective feature amount can be selected from a plurality of feature amounts mixed with noise.

  According to the present invention, since an effective feature amount can be selected from a plurality of feature amounts mixed with noise, it is possible to correctly determine a document unit based on these feature amounts. In addition, since the separation determination can be automated, the burden on the user in the manual separation work can be reduced, and efficient separation work can be performed.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

  FIG. 1 is a block diagram illustrating a configuration example of a document filing apparatus according to an embodiment of the present invention. As is apparent from FIG. 1, the document filing apparatus according to the present embodiment includes a document input unit 11, a feature amount extraction unit 12, a similarity evaluation unit 13, a document storage unit 14, a document delimiter unit 15, and a document output unit 16. These components are connected to each other via a bus line 17. In the document filing apparatus having such a configuration, the feature quantity extraction unit 12 is the most characteristic part of the present invention.

  The document input unit 11 acquires document data from the input document and inputs it as document data to be registered in the document filing apparatus. Here, the input document data includes, for example, multi-value image data obtained by scanning from a printed material, or multi-value image data captured by a digital camera. Correspondingly, as the document input unit 11, for example, a scanner device provided with an ADF (Auto Document Feeder) and its control means, or a memory for continuously extracting images stored in a memory (card) of a digital camera A reader device and its control means are used.

  The feature quantity extraction unit 12 extracts a quantity (hereinafter referred to as “feature quantity”) indicating a characteristic unique to the document data from the multivalued document data (multivalued image data) input by the document input unit 11. To do. In the feature quantity extraction unit 12 according to this example, as shown in FIG. 2, a projection waveform is formed by accumulating pixel values of each document data in the horizontal and vertical directions, and the height of a character representing the document data is determined from the projection waveform. For example, n pieces of height h and width w are extracted. Here, the projection waveform data is the integrated value or average value of the vertical and horizontal pixel values of each document data. For example, the projection data when the vertical or horizontal direction of each document data is all black. The value of "255" is "0" when all are white.

  FIG. 3 is a block diagram illustrating an example of a specific configuration of the feature amount extraction unit 12. As is clear from FIG. 3, the feature quantity extraction unit 12 according to this example has a configuration including a projection waveform generation unit 21, a binarization unit 22, and a frequency distribution analysis unit 23.

  The projection waveform generation unit 21 forms a projection waveform directly from the input multivalued document data. Specifically, as described above, the projection waveform data is generated by accumulating the pixel values of each document data in the horizontal and vertical directions (or taking the average). The binarization unit 22 performs binarization processing on the projection waveform data generated by the projection waveform generation unit 21 and outputs a binary data series consisting of only a black region and a white region. The frequency distribution analysis unit 23 obtains all the lengths of the continuous black areas in the binary data series output from the binarization unit 22, and for example, as shown in FIG. N / 2 pieces higher than a predetermined threshold are set as the feature amount of the document data together with the appearance frequency.

  The frequency distribution analysis unit 23 obtains the length of the continuous white area in the binary data series. For example, those values whose appearance frequency is higher than a predetermined threshold are included in the document together with the appearance frequency. You may make it include in the feature-value of data. That is, all the values indicating the characteristics of at least one of the continuous black area and white area in the binarization result are obtained and the frequency distribution is analyzed, and the analysis result is used as the feature amount of the document data.

  Further, for example, a representative value such as an average value, a median value, or a quartile of the obtained values may be obtained, and the representative value may be a part of the feature amount of the document data. Further, as shown in FIG. 5, the projection waveform data itself or its binarization result is divided, the representative value as described above is obtained from each of the divided pieces, and the representative value is used as one of the feature values of the document data. It is good as a part. However, the method for obtaining the feature value is not limited to the above method.

  Referring back to FIG. 1, the document storage unit 14 stores and stores the input document data in association with the n types of feature values extracted by the feature value extraction unit 12, and includes a hard disk drive and a DVD (Digital Versatile Disc). -Realized by a mass storage device such as a RAM / ± RW / ± R drive.

  The similarity evaluation unit 13 extracts a plurality of feature amounts if a plurality of feature amounts are accumulated for the feature amounts extracted by the feature amount extraction unit 12 and associated with the document data and accumulated in the document accumulation unit 14. Are respectively compared to obtain the similarity between them. The similarity referred to here is, for example, the Euclidean distance between feature vectors associated with each document data if the feature value is a vector expression (hereinafter, this feature value is referred to as “feature vector”). It is a degree evaluated based on.

  Specifically, as shown in FIG. 6A, the case where the Euclidean distance between feature vectors is smaller than a predetermined threshold is evaluated as similarity: large, and as shown in FIG. 6B. The case where the Euclidean distance between feature vectors is equal to or greater than the predetermined threshold is evaluated as similarity: small. However, the definition of the distance between feature vectors is not limited to the Euclidean distance.

  The document delimiter 15 obtains a similarity between a series of document data by the similarity evaluation unit 13 and puts a delimiter into the series of document data based on the obtained similarity. As a method of accumulating in document units (consisting of document data of one page or more), for example, a method of creating a file folder for each document and storing corresponding document data with a file name having a sequential number in the input order Alternatively, there is a method of using an image file format such as a multi-page TIFF (Tagged Image File Format) that can hold a plurality of pages.

  The document output unit 16 outputs document data instructed to be output in a predetermined format. For example, a CRT (Cathode Ray Tube) and its control means, a printer device and its control means, a magnetic disk, a memory card, etc. The read / write device and its control means, or a data transfer device that exchanges data via a network or the like. That is, the document output unit 17 outputs, for example, a document printed on paper, image data output to a CRT, or a file formatted by HTML (Hyper Text Markup Language) as an output result.

  Next, the procedure of document data separation processing in the document filing apparatus according to the present embodiment having the above-described configuration will be described with reference to the flowchart of FIG.

  For example, the user sets a paper document including one page or more on the ADF. At this time, the paper document may be either a single document (consisting of one or more pages) or a plurality of documents (comprising one or more pages). Also, when setting a paper document, the user need not be aware of document separation. The paper document set in the ADF is sent to the scanner device page by page by the ADF. At this time, the scanner device functions as the document input unit 11 in FIG. That is, from the document input unit 11, the same number of document data as the number of pages of the paper document set in the ADF is input to the document filing device.

  When document data is input from the document input unit 11 (step S11), whether the input document data has already been extracted with a feature amount and is stored in the document storage unit 14 in association with the feature amount. It is determined whether or not (step S12). If the input document data is document data for which feature amounts have not yet been extracted, the feature amount is extracted from the input document data by the feature amount extraction unit 12 (step S13), and then the process proceeds to step S14. To do. If the input document data is already extracted document data, the process directly proceeds to step S14. That is, the process for extracting the feature amount is not performed on the document data in which the feature amount has already been extracted from the input document data and is stored in the document storage unit 14 in association with the feature amount.

  The above processing is performed for all input document data (step S14). Subsequently, the feature amount associated with the document data stored in the document storage unit 14 and the feature amount associated with the input document data are compared with each other with respect to the document data stored in the document storage unit 14. Evaluation is performed by the evaluation unit 13 (step S15). Then, based on the evaluation result, a document unit separation result is determined for the document data, separated into document units, and stored in the document storage unit 14 (step S16).

  In the above-described embodiment, when the feature amount extraction unit 12 extracts feature amounts from the binary data series, for example, all the lengths of continuous black regions in the binary data series are obtained, and the values of those values are obtained. Of these, n / 2, whose appearance frequency is higher than a predetermined threshold, is used as the feature amount of the document data together with the appearance frequency. However, the present invention is not limited to this, and the following method can also be adopted. It is.

  When extracting a feature amount from a binary data series, for example, the skewness is obtained from values obtained by obtaining all the lengths of continuous black areas in the binary data series, and the skewness is obtained from the document data. It may be a part of the feature amount. If this document data is printed as plain text, it is expected that the document data has a symmetrical distribution as shown in FIG. 8, and the obtained skewness is close to zero.

  Further, when extracting feature values from a binary data series, for example, kurtosis is obtained from values obtained by obtaining all the lengths of continuous black regions in the binary data series, and the kurtosis is obtained from this document. It may be a part of the data feature. If this document data is printed as plain text, the value is expected to fluctuate within a narrow range as shown in FIG. Become.

It is a block diagram which shows the structural example of the document filing apparatus which concerns on one Embodiment of this invention. It is a conceptual diagram of the feature-value extraction in a feature-value extraction part. It is a block diagram which shows an example of the specific structure of a feature-value extraction part. It is explanatory drawing of the specific example which extracts the feature-value from a binary data series. It is explanatory drawing of the other specific example which extracts the feature-value from a binary data series. It is a conceptual diagram of similarity evaluation. It is a flowchart which shows the procedure of the division | segmentation process of document data. It is explanatory drawing of the other example which extracts the feature-value from a binary data series. It is explanatory drawing of the further another example which extracts the feature-value from a binary data series. It is FIG. (1) with which it uses for description of the subject of a prior art. It is FIG. (The 2) with which it uses for description of the subject of a prior art.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 11 ... Document input part, 12 ... Feature quantity extraction part, 13 ... Similarity evaluation part, 14 ... Document storage part, 15 ... Document delimiter part, 16 ... Document output part, 21 ... Projection waveform generation part, 22 ... Binarization Part, 23 ... frequency distribution analysis part

Claims (12)

Projection waveform generation means for generating projection waveform data from input multi-value document data;
Binarization means for binarizing the projection waveform data generated by the projection waveform generation means;
The frequency distribution is analyzed for all the values indicating the characteristics of at least one of the continuous black area and the white area in the binarization result by the binarization means, and the analysis result is analyzed for the characteristics of the document data. A feature quantity extraction device comprising: an analysis means for obtaining a quantity. The analysis means obtains a length of a continuous black region in the binarization result, and includes, among the obtained values, an appearance frequency higher than a predetermined threshold value in the feature amount. The feature amount extraction apparatus according to claim 1. The analysis means obtains the length of a continuous white region in the binarization result, and includes, in the feature amount, an appearance frequency higher than a predetermined threshold value among the obtained values. The feature amount extraction apparatus according to claim 1. The feature amount extraction apparatus according to claim 1, wherein the analysis unit sets a representative value of the obtained value as a part of the feature amount. The feature value extraction apparatus according to claim 4, wherein the representative value is an average value, a median value, or a quartile of the obtained values. The feature amount extraction apparatus according to claim 4, wherein the analysis unit divides the projection waveform data or the binarization result and obtains the representative value from each of the divided parts. The feature amount according to claim 1, wherein the analysis unit obtains a skewness from values obtained for all the lengths of the continuous black regions, and the skewness is a part of the feature amount. Extraction device. The feature amount according to claim 1, wherein the analysis unit obtains a kurtosis from values obtained for all the lengths of the continuous black regions, and uses the kurtosis as a part of the feature amount. Extraction device. A first step of generating projection waveform data from input multivalued document data;
A second step of binarizing the projection waveform data generated in the first step;
The frequency distribution is analyzed by obtaining all the values indicating the characteristics of at least one of the continuous black area and the white area in the binarization result in the second step, and the analysis result is analyzed for the characteristics of the document data. A feature amount extraction method comprising: a third step of making a quantity. In the third step, the length of the continuous black region in the binarization result is obtained, and the obtained value includes a value having an appearance frequency higher than a predetermined threshold among the obtained values. The feature quantity extraction method according to claim 9. Projection waveform generation means for generating projection waveform data from input multi-value document data;
Binarization means for binarizing the projection waveform data generated by the projection waveform generation means;
The frequency distribution is analyzed for all the values indicating the characteristics of at least one of the continuous black area and the white area in the binarization result by the binarization means, and the analysis result is analyzed for the characteristics of the document data. An analysis means for quantity;
A document filing apparatus comprising: a determination unit that determines a break of a document unit based on the feature amount obtained by the analysis unit. The analysis means obtains a length of a continuous black region in the binarization result, and includes, among the obtained values, an appearance frequency higher than a predetermined threshold value in the feature amount. The document filing apparatus according to claim 11.