JP5051971B2 - Image encoding device, image decoding device, image encoding method, computer-readable recording medium recording program of image encoding method, and computer-readable recording medium recording program of image decoding method and image decoding method - Google Patents

Description

  The present invention relates to an image encoding device, an image encoding method, a computer-readable recording medium on which a program for the image encoding method is recorded, and a computer-readable recording medium on which an image decoding method and a program for the image decoding method are recorded. . In particular, an image encoding apparatus, an image encoding method, and a computer-readable recording medium on which a program for the image encoding method is recorded, and image decoding, in which an input pattern replacement error due to encoding is reduced while maintaining high encoding efficiency The present invention relates to a computer-readable recording medium on which a program for an image decoding method and an image decoding method are recorded.

  Conventionally, as a typical method of encoding a document image, a method of encoding a character part as a character code using a character recognition technology, and a method of encoding in the same way as encoding normal image data, There is.

  The method using the character recognition technique has a feature that the data capacity after encoding becomes small. However, although performance is improved, character recognition errors cannot be made zero. For this reason, if erroneous encoding is performed, it may occur that the understanding of the text included in the image is hindered at the time of decoding.

  On the other hand, a method of encoding in the same way as a normal image applies a generally well-known image compression method to a document image as it is. In this method, unless the image quality is extremely deteriorated, it is difficult to cause trouble in understanding the sentence. However, the data capacity after encoding becomes larger than in the case of using character recognition technology.

  In order to cover the disadvantages of both, there is a method in which a plurality of similar patterns are represented by a single pattern and only the representative pattern, the identification code of the representative pattern, and the appearance position of the pattern represented by the representative pattern are encoded. It has been proposed before. This encoding method is described in “RNAscher et al.“ A Means for Achieving a High Degree of Compaction on Scan-Digitized Printed Text ”, IEEE Transactions on Computers, vol.c-23, No.11, November 1974” ( Non-patent document 1) and the like are disclosed in detail.

  The pattern here corresponds to a character in many cases in encoding of a document image. For this reason, apart from the data capacity of the encoded data of the representative pattern itself, the data capacity of the encoded data ideally represents the identification code of each character in the image and the corresponding position information. The amount of data required will be sufficient.

  This method can also be considered as a standard pattern extracted from an input image in a character recognition technique using a pattern matching method.

  In this method, recognition errors are not a problem as much as when character recognition technology is used. This is because, in this method, it is only necessary to determine whether or not the patterns are similar, and it is not necessary to accurately determine what the pattern is really like as in character recognition. When compared with character recognition using a pattern matching method, a pattern corresponding to a standard pattern is extracted from the input image itself. For this reason, even if a character of a special font is included in the input image, this does not hinder the encoding itself.

  As described above, the method of encoding the representative pattern has relatively excellent properties. Nevertheless, it is not widely used at present.

  This is because encoding is difficult. This is because it is difficult to control so as to reduce an input pattern replacement error, that is, an error that replaces an input pattern with an incorrect representative pattern.

  An example of an input pattern replacement error will be described with reference to FIG. FIG. 46A shows an input image, and FIG. 46B shows an image after the input image is encoded and decoded. In the input image of FIG. 46 (a), there are three sets of similar patterns, “su” and “nu”, “sawa” and “choice”, and “po” and “bo”. In such a case, an input pattern replacement error is likely to occur, and as shown in FIG. 46B, “su”, “choice”, and “po” should not be replaced by “nu”, “ It has been replaced with a pattern representing “sawa” and “bo”. This occurs because the input pattern clustering is inappropriate.

  47 to 52, each input pattern (character) is expressed as a point on a two-dimensional plane. This figure does not show the position of the input pattern on the input image, but schematically shows the position of the feature vector created by extracting the features of the input pattern in the pattern space (feature vector space). Yes. 47 to 52, the feature vector is expressed as a point on a two-dimensional plane. However, when there are three or more feature amounts, the pattern space has three or more dimensions.

  Referring to FIG. 47, input patterns representing two types of characters are represented by pattern 102 (“◯” mark) and pattern 104 (“Δ” mark), respectively.

  Here, the representative pattern is selected from the input patterns. For example, input patterns whose Euclidean distance is within a certain range in the pattern space are classified into one class, and a representative pattern representing that class is selected. For example, referring to FIG. 48, patterns 102 and 104 are classified into classes represented by three circles 112, 114 and 116, and patterns 106, 108 and 110 are selected as representative patterns representing the three classes, respectively. It is. Note that the representative pattern extraction method is not limited to the clustering method based on the Euclidean distance. Correct replacement of input patterns means that each representative pattern is selected from input patterns representing the same character. FIG. 48 shows an example of ideal clustering in which correct replacement of input patterns is performed.

  The three circles 112, 114, and 116 are circles having a constant radius centered on the patterns 106, 108, and 110, respectively. The pattern in the circle is replaced with the representative pattern at the time of encoding. At this time, the same type of input pattern is not necessarily represented by one representative pattern. As shown in FIG. 48, the pattern 102 is contained within two circles 112 and 114. For this reason, the pattern 102 is represented by two patterns 106 and 108.

  When the diameter of the circle, that is, the Euclidean distance between the representative pattern considered to belong to the same class and the input pattern is increased, all the input patterns are included in the circle 118 as shown in FIG. 49 and clustered into one class. Will be. For this reason, different types of patterns 102 and 104 are represented by the same representative pattern 120, and an input pattern replacement error as shown in FIG. 46B occurs.

  As described above, if the diameter of the circle in clustering is increased, an input pattern replacement error is likely to occur, and if the circle diameter is decreased, a replacement error is less likely to occur. For this reason, it seems to be sufficient to reduce the diameter of the circle.

  However, as shown in FIG. 50, when the diameter of the circle is made as close to 0 as possible, the replacement pattern does not occur, but the input pattern and the representative pattern have a one-to-one correspondence. For this reason, even if the input pattern is encoded using the representative pattern, it is not different from the case where the input pattern itself is encoded, and the data amount is not reduced.

  Thus, there is a trade-off relationship between the reduction of the data amount and the reduction of the input pattern replacement error.

  Japanese Patent Laid-Open No. 8-30794 (Patent Document 1) discloses a method of extracting a representative pattern from an input pattern as described below. That is, a pattern called a registered pattern is prepared in addition to the input pattern and the representative pattern. First, one registered pattern is selected from the input patterns, and the registered pattern and the input pattern are sequentially verified. If the registered pattern is similar to the input pattern, the pattern obtained by calculating the average of the registered pattern and the input pattern, or selected from the registered pattern or the input pattern based on a predetermined criterion The pattern is a new registered pattern. Input patterns similar to registered patterns are clustered into the same class.

When an input pattern that is not similar to any registered pattern occurs, the input pattern is set as a new registered pattern, and the same processing is performed. Such a process is performed until the input pattern is clustered into any class, and the registered pattern at the end of the process is set as the representative pattern.
JP-A-8-30794 RNAscher et al. "A Means for Achieving a High Degree of Compaction on Scan-Digitized Printed Text", IEEE Transactions on Computers, vol.c-23, No. 11, November 1974

  However, even if such a representative pattern registration method is used, the input pattern clustering method is the same as the above-described method. For this reason, it is difficult to achieve both a reduction in data amount and a reduction in input pattern replacement error.

  For example, referring to FIG. 51, when an input pattern 201 is registered as a registered pattern and an attempt is made to perform clustering so that an input pattern replacement error does not occur, the input pattern 202 belongs to the same class, but the input pattern 203 belongs to another class. Will belong. For this reason, the number of representative patterns increases.

  Referring to FIG. 52, conversely, if pattern 102 is to be represented by one representative pattern, input pattern 204 of pattern 104 belongs to the same class, and an input pattern replacement error occurs.

  In addition, special attention is required for processing a document image of a language including many characters composed of a plurality of connected components such as Japanese. In the above-mentioned Japanese Patent Application Laid-Open No. 8-30794, a connected component is used as an input pattern, and it is not considered whether the input pattern is extracted from the same character or from different characters. Therefore, it is possible that a plurality of connected components constituting so-called separation characters are replaced with representative patterns extracted from different characters at the time of encoding on the decoded image. However, when the typeface of the character that is the basis of such a representative pattern is different, a noticeable discomfort is caused on the decoded image. For example, referring to FIG. 53, the “body” prejudice (left part) written in the Mincho style and the “forest” fence (right part) written in Gothic are used as representative patterns, respectively. Suppose that a Gothic “rest” on the same sheet is encoded. In this case, in the decoded image, the “vacation” is taken from a typeface that has a partial portion in the Mincho style and a heel portion in the Gothic style, and is uncomfortable. In order to prevent this, conventionally, there is only a strict condition for the representative pattern to replace the input pattern. This increases the number of representative patterns as described above, and lowers the encoding efficiency.

  The present invention has been made to solve the above-described problems, and an object of the present invention is to provide an image encoding apparatus that reduces input pattern replacement errors due to encoding while maintaining high encoding efficiency. .

  Another object of the present invention is to prevent a sense of incongruity between encoded characters of separated characters while maintaining encoding efficiency.

  In order to solve the above-described problems, according to one aspect of the present invention, an image encoding device includes an input pattern extraction unit for extracting an input pattern from image data, and input patterns extracted by the input pattern extraction unit. For each part constituting the input pattern, and representative pattern extraction means for extracting the representative pattern from among the input patterns similar to each other, and for outputting the image of the representative pattern and the coordinate position of the input pattern Output means.

  According to the image coding apparatus having the above-described configuration, by comparing input patterns partially, it is possible to distinguish characters that are similar on the whole but not similar on a partial basis. . For this reason, the replacement error of the input pattern can be reduced.

  According to another aspect of the present invention, an image encoding device is similar to an input pattern extraction unit for extracting an input pattern from image data and the input pattern extracted by the input pattern extraction unit. A similar enlarging means for making an input pattern similar to the input pattern to be an input pattern similar to the input pattern, and a representative pattern to be extracted from the input patterns determined to be similar to each other by the similar enlarging means Representative pattern extraction means, and output means for outputting the representative pattern image and the input pattern coordinate position.

  According to the image coding apparatus having the above-described configuration, the number of representative patterns representing an input pattern can be reduced by expanding the similarity range of input patterns in a chain. For this reason, encoding efficiency can be kept high.

  According to another aspect of the present invention, an image encoding method includes a step of extracting an input pattern from image data, and the extracted input patterns are compared with each other constituting the input pattern, and inputs similar to each other. A step of extracting a representative pattern from the pattern, and a step of outputting the image of the representative pattern and the coordinate position of the input pattern.

  According to the image coding method having the above-described configuration, by comparing input patterns partially, it is possible to distinguish characters that are similar on the whole but not similar on a partial basis. . For this reason, the replacement error of the input pattern can be reduced.

  According to another aspect of the present invention, an image encoding method compares an extracted input pattern with a step of extracting an input pattern from image data, and an input pattern similar to the input pattern for each of the input patterns. The input pattern similar to the input pattern is set to the input pattern similar to the input pattern, the representative pattern is extracted from the input patterns similar to each other, and the image of the representative pattern and the coordinate position of the input pattern are output. Including the step of.

  According to another aspect of the present invention, an image decoding method decodes an image from data encoded by the above image encoding method. The image decoding method includes a step of extracting a representative pattern image and an input pattern coordinate position from encoded data, and a step of pasting a representative pattern representing the input pattern to the input pattern coordinate position.

  According to the image decoding method having the above configuration, an image can be created simply by sequentially pasting the representative pattern to the coordinate position of the input pattern. For this reason, an image can be restored at high speed.

  According to another aspect of the present invention, the recording medium is a computer-readable recording medium that records a computer-executable image encoding method program. In the image encoding method, the step of extracting an input pattern from image data, the extracted input patterns are compared for each part constituting the input pattern, and a representative pattern is extracted from mutually similar input patterns. And a step of outputting an image of the representative pattern and a coordinate position of the input pattern.

  With the above recording medium, the computer can distinguish characters that are similar in overall view but not similar in partial view by partially comparing the input patterns. For this reason, the replacement error of the input pattern can be reduced.

  According to another aspect of the present invention, the recording medium is a computer-readable recording medium that records a computer-executable image encoding method program. In the image encoding method, an input pattern is extracted from image data, and the extracted input patterns are compared with each other. For each input pattern, an input pattern similar to the input pattern is input to the input pattern. An input pattern similar to the pattern is included, a representative pattern is extracted from among the input patterns similar to each other, and an image of the representative pattern and a coordinate position of the input pattern are output.

  According to still another aspect of the present invention, the recording medium is a computer-readable recording medium that records a computer-executable image decoding method program. The image decoding method is a method of decoding an image from the data encoded by the above image encoding method. The image decoding method includes a step of extracting a representative pattern image and an input pattern coordinate position from encoded data, and a step of pasting a representative pattern representing the input pattern to the input pattern coordinate position.

  According to the above recording medium, the computer can create an image simply by sequentially pasting the representative pattern to the coordinate position of the input pattern. For this reason, an image can be restored at high speed.

  By comparing the input patterns partially, it is possible to distinguish characters that are similar in whole but not similar in partial. For this reason, the replacement error of the input pattern can be reduced.

  Also, by detecting the number of loops, although partially similar, it is possible to accurately distinguish different characters. For this reason, the replacement error of the input pattern can be reduced.

  Furthermore, the number of representative patterns representing an input pattern can be reduced by chaining the similarity range of the input pattern. For this reason, encoding efficiency can be kept high.

  Furthermore, characters cut out from the image data are used as the representative pattern. Therefore, the input pattern replacement error due to character recognition does not occur as in the case where the input pattern is recognized and the representative pattern is represented by the character code.

  At the time of image decoding, an image can be created by simply pasting the representative pattern to the coordinate position of the input pattern. For this reason, an image can be restored at high speed.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the following description, the same parts are denoted by the same reference numerals. Their names and functions are also the same. Therefore, detailed description thereof will not be repeated.

  Referring to FIG. 1, an image encoding apparatus according to an embodiment of the present invention is connected to scanner 303 that scans a paper surface and captures an image, and automatically feeds the paper surface sequentially to scanner 303. An auto feeder 301, a counter 302 that is connected to the auto feeder 301 and counts the number of pages of paper fed to the scanner 303, and image data that is connected to the scanner 303 and stores an image captured by the scanner 303 Buffer 304.

  The image encoding device is further connected to an image data buffer 304 and connected to a binarization threshold value calculator 307 for calculating a binarization threshold value for each page, and a binarization threshold value calculator 307. Input pattern extraction for extracting an input pattern from an image, connected to a binarization threshold buffer 308 that stores a binarization threshold as a one-dimensional array for each page, and an image data buffer 304 and a binarization threshold buffer 308 305.

  The image encoding apparatus is further connected to an input pattern extractor 305, and a page counter 306 that counts the number of pages of an image currently being processed, an input pattern image buffer 309 that stores an image of the input pattern, and an input pattern extraction An input pattern information buffer 310 for storing the horizontal and vertical widths of the input pattern, and a representative pattern for extracting a representative pattern representative of the input pattern, connected to the input pattern image buffer 309 and the input pattern information buffer 310. And an extractor 311.

  The image coding apparatus is further connected to a representative pattern extractor 311, an input pattern image buffer 309, and an input pattern information buffer 310, and a representative pattern label buffer 312 that stores an integer array for associating the representative pattern with the input pattern The representative pattern image buffer 313 that is connected to the representative pattern extractor 311 and the representative pattern label buffer 312 and stores the image of the representative pattern, and the representative that is connected to the representative pattern extractor 311 and stores the horizontal width and the vertical width of the representative pattern. Pattern information buffer 314.

  The image encoding apparatus is further connected to the representative pattern information buffer 314, and the representative pattern information compressor 315 for compressing the representative pattern and the representative pattern image buffer 313 are connected to the representative pattern image buffer 313 and stored in the representative pattern image buffer 313. Connected to a representative pattern image color reducer 316 that reduces the color of the image, an input pattern information buffer 310 and a representative pattern label buffer 312, information on the number of pages stored in the counter 302, and information stored in the input pattern information buffer 310 And an input pattern information compressor 317 for mixing and compressing information stored in the representative pattern label buffer 312.

  The image encoding apparatus is further connected to the representative pattern information compressor 315 and the input pattern information compressor 317, and connects the representative pattern information, the representative pattern image, and the compressed data of the input pattern information to one encoded data. A representative pattern image that is connected to the representative pattern image color reducer 316 and connected to the representative pattern image color reducer 316 and compresses the representative pattern reduced by the representative pattern image color reducer 316. A compressor 318 and an encoded data buffer 320 connected to the data mixer 319 and storing data obtained by encoding the document image are included.

  Referring to FIG. 2, input pattern extractor 305 is connected to image data buffer 304, and includes a character element extraction unit 701 that extracts character elements from an image stored in image data buffer 304, and character element extraction unit 701. A character element buffer 702 that stores the character elements extracted by the character element extraction unit 701, a character string direction determination unit 703 that is connected to the character element buffer 702 and determines the direction of the character string in the image, and a character It includes a character string direction information flag 713 that is connected to the column direction determination unit 703 and stores the direction of the character string.

  The input pattern extractor 305 is further connected to the character element buffer 702 and the character string direction information flag 713, and is connected to the character string extraction unit 705 for extracting a character string from the image and the character string extraction unit 705 for extraction. Connected to a character string information buffer 706 that stores an integer array in which character string numbers and character elements are associated one-to-one, a character element buffer 702, a character string extraction unit 705, and a character string information buffer 706, And an individual character extraction unit 707 that divides the character string into character candidates.

  The input pattern extractor 305 is further connected to the individual character extraction unit 707, and includes an individual character information buffer 708 for storing coordinates of circumscribed rectangles of character candidates, a character string counter 709 for counting the number of character strings, and a character Counter 710 that counts the number of characters, character counter 711 that counts the number of characters in a certain character string, binarization threshold buffer 308, individual character extraction unit 707, character string counter 709, character counter 710, connected to the individual character information buffer 708, the character counter in character string 711, the binarization threshold buffer 308 and the binarization threshold calculator 307, and compares the character standard pattern 712 and the character extracted from the character string And a character matching unit 704 that performs.

  Referring to FIG. 3, representative pattern extractor 311 includes a loop detector 1001 that detects the number of loop-shaped portions (loops) included in the image corresponding to the input pattern, and the number of loops that stores the number of loops. A buffer 1002, a first counter 1003 that counts the number of input patterns, a second counter 1004 that counts the number of input patterns, a pattern comparator 1005 that compares input patterns with each other, a loop detector 1001, and a loop number buffer 1002, a first counter 1003, a second counter 1004, and a pattern comparator 1005, and a controller 1000 that controls the connected devices.

  Referring to FIG. 4, a loop detector 1001 includes a first counter 1301 indicating the number of an input pattern currently being processed, a connected component circumscribed rectangle extractor 1302 that extracts a circumscribed rectangle of a connected component included in the input pattern, and A connected component circumscribed rectangle information buffer 1303 for storing the coordinates of the circumscribed rectangle of the connected component, a second counter 1304 for counting the number of connected components, a first counter 1301, a connected component circumscribed rectangle extractor 1302, a connected component circumscribed A controller 1300 is connected to the rectangular information buffer 1303 and the second counter 1304, and controls the connected devices.

  Referring to FIG. 5, a pattern comparator 1005 extracts a feature amount from an input pattern and converts it into a feature vector, a vector normalizer 1602 that normalizes the feature vector, a feature vector A vector canonicalizer 1603 that performs canonicalization, an inner product calculator 1604 that performs inner product calculation of feature vectors, a counter 1605 that counts the number of partial vectors, and a partial vector generator 1606 that generates partial vectors from feature vectors And a controller 1600 connected to the vector converter 1601, the vector normalizer 1602, the vector canonicalizer 1603, the inner product calculator 1604, the counter 1605, and the partial vector generator 1606 and controlling the connected devices.

  In the image encoding apparatus, it is assumed that a plurality of sheets are input by a scanner with an auto feeder, but the present invention is not limited to this.

  Referring to FIG. 6, an image decoding apparatus that decodes encoded data created by the image encoding apparatus into an image is connected to encoded data buffer 2201 that stores the encoded data, and encoded data buffer 2201. A data separator 2202 that separates the encoded data into representative pattern information, representative pattern image, and input pattern information; and a representative pattern information decompressor 2203 that is connected to the data separator 2202 and decompresses the representative pattern information. Including.

  The image decoding apparatus is further connected to a representative pattern information decompressor 2203 and is connected to a representative pattern information buffer 2206 for storing the decompressed representative pattern information and a data separator 2202, and a representative pattern image for decompressing the representative pattern image. Connected to the decompressor 2204, the representative pattern image decompressor 2204, and the representative pattern image buffer 2207 for storing the decompressed representative pattern image, and the input pattern connected to the data separator 2202 for storing the compressed input pattern information It includes a compression information buffer 2205 and a pixel value conversion table 2209 in which a table for converting pixel values is stored.

  The image decoding apparatus is further connected to the representative pattern image buffer 2207 and the pixel value conversion table 2209, and is connected to the representative pattern pixel value converter 2208 for converting the pixel value of the representative pattern and the representative pattern information buffer 2206, and the representative pattern. Based on the data stored in the information buffer 2206, the representative pattern image offset generator 2210 for generating the storage position of the representative pattern image in the representative pattern image buffer 2207, and the representative pattern image offset generator 2210 are connected. A representative pattern image offset table 2211 that stores an integer array in which numbers and offset values are associated with each other.

  The image decoding apparatus is further connected to an input pattern compression information buffer 2205, and an input pattern information offset generator 2212 that generates an offset representing the position of each page data in the input pattern compression information buffer 2205, and an input pattern information offset Connected to the generator 2212, connected to the input pattern information offset table 2213 for storing an integer array in which page numbers and offsets are associated, a page counter 2214 for counting the number of pages, and an input pattern compression information buffer 2205. And an input pattern information decompressor 2217 for decompressing the input pattern information.

  The image decoding apparatus is further connected to an input pattern information decompressor 2217, stores an input pattern information buffer 2218 for storing input pattern information, an input pattern counter 2219 for counting the number of input patterns, and stores an image for each page. A page image buffer 2215 and a page image buffer initializer 2216 that is connected to the page image buffer 2215 and initializes pixel values of an image stored in the page image buffer 2215 are included.

  The image decoding apparatus is further connected to a page image buffer 2215, and displays a display device 2221 that displays an image stored in the page image buffer 2215, a representative pattern image offset table 2211, an input pattern information buffer 2218, an input pattern counter 2219, Pixel density conversion that is connected to the page image buffer 2215, the representative pattern image buffer 2207, the representative pattern pixel value converter 2208, the pixel value conversion table 2209, and the input pattern information offset table 2213, and makes the size of the representative pattern equal to the size of the input pattern Instrument 2220.

  The image encoding process will be described in detail with reference to FIG. Hereinafter, the expression “page” is used to indicate each page of the document. The element number or page number of the array starts from 0 unless otherwise specified. Furthermore, the loop variables i, j, and k are repeatedly used to describe the operation of different parts, but the operations are irrelevant unless otherwise specified.

  The auto feeder 301 clears the counter 302 to 0 (step S (hereinafter, “step” is omitted) 401). When the counter value is i, the scanner 303 scans the i-th page and stores the captured image in the image data buffer 304 (S402). An example of data stored in the image data buffer 304 is shown in FIG.

  The image stored in the image data buffer 304 is a grayscale image with 256 gradations in which one pixel is represented by one byte. The auto feeder 301 increments the counter 302 by 1 (S403). If the value of the counter 302 is not equal to the number of pages (NO in S404), it means that there is an unscanned page (i-th page). Therefore, the process after S402 is repeated.

  When the value of the counter 302 becomes equal to the number of pages (YES in S404), the input pattern extractor 305 clears the page counter 306 to 0 (S405). The binarization threshold value calculator 307 takes out the image of the page pointed to by the page counter 306 from the image data buffer 304, calculates the optimum binary threshold value, and stores it in the binarization threshold value buffer 308 (S406). . The binarized threshold value is used for input pattern extraction processing described later. The binarization threshold value calculator 307 determines the binarization threshold value so that the variance between the character area and the background area (so-called intergroup variance) is maximized for each page image. The binarization threshold value calculation method using intergroup variance is disclosed in detail, for example, on page 43 of Nagao “Image Recognition Theory” (Corona, 1983). Note that the binarization threshold value calculation method is not limited to this. If the binarized threshold value is not necessary for the input pattern extraction process, S406 may be omitted.

The binarization threshold buffer 308 has an array corresponding to each page on a one-to-one basis, and stores an optimum binarization threshold for each page. Here, a pixel having a pixel value smaller than the binarization threshold is set as a non-background pixel, that is, a target pixel. If the i-th element of the array TH stored in the binarization threshold buffer 308 is TH [i], the pixel of interest included in the i-th page is a pixel that satisfies Expression (1).
0 ≦ pixel value <TH [i] (1)
Here, the image stored in the image data buffer 304 is a grayscale image having 256 gradations, but a processing method in other cases will be additionally described. When the input image is color, for example, a binarization threshold value is calculated only for the luminance component.

  If the input image is a binary image, the distinction between non-background pixels and background pixels is self-evident even if threshold processing is not performed, and therefore S406 can be omitted.

  The input pattern extractor 305 performs binarization processing of the page image indicated by the page counter 306 based on the equation (1). Thereafter, an input pattern is extracted from the binary image (S407). Here, the input pattern refers to a small area that is considered to have many similar patterns in the paper.

  Referring to FIG. 9, when a part of the input image is enlarged, it is assumed that “Akai River” and a figure that is not a character but resembles a human face having a size almost similar to the character exists. . When a connected component of black pixels (character area pixels) is obtained from this portion and an input pattern is extracted, it is as shown in FIG. That is, twelve input patterns are obtained from the character string “Akaigawa” and the following graphic.

  When characters are cut out from the same part and an input pattern is obtained, it is as shown in FIG. That is, five input patterns are obtained from the character string “Akaigawa” and the graphic following it. In the following description, as shown in FIG. 11, a character is described as an input pattern, but is not limited to a character.

  The input pattern extractor 305 increments the page counter 306 by 1 (S408). If the value indicated by the page counter 306 does not match the number of pages (NO in S409), the processes in and after S406 are repeated until input patterns are extracted for all pages.

  If the value indicated by the page counter 306 matches the number of pages (YES in S409), the representative pattern extractor 311 refers to the input pattern image buffer 309 and the input pattern information buffer 310 to extract a representative pattern, The result is stored in the representative pattern label buffer 312 and the representative pattern information buffer 314 (S410). The representative pattern here means a pattern that can replace the input pattern in the input image without greatly degrading the image quality. The process of S410 will be described in detail later.

  The representative pattern information compressor 315 compresses the representative pattern stored in the representative pattern information buffer 314 (S412). The representative pattern image color reducer 316 reduces the color of the representative pattern image stored in the representative pattern image buffer 313 (S413). The reason for color reduction is to further reduce the amount of information to the number of gradations necessary to reproduce a character pattern that is expected to occupy most of the input pattern, and to further improve the compression rate. Here, it is assumed that the representative pattern of 256 gradations is reduced to 8 gradations. 256 gradations are divided into eight equal parts, and the eight representative colors 0, 36, 73, 109, 145, 181, 218, and 255 are examined, and the closest representative color is assigned first. The pixel value is replaced with a number indicating. For example, the pixel value 120 is closest to the representative color 109, and the representative color 109 corresponds to the third representative color, with the representative color 0 being 0th when numbers are assigned to the representative colors in ascending order. Therefore, the pixel value 120 is replaced with 3 in S413.

  The representative pattern image compressor 318 compresses the representative pattern reduced by the representative pattern image color reducer 316, and supplies it to the data mixer 319 (S414). The compression method is roughly divided into a method of compressing a two-dimensional structure as an image, and a method of compressing as a one-dimensional simple array, and any method can be used for compression. Here, compression is performed as a one-dimensional array, and an entropy coding method using arithmetic coding is used. The data compression method is not limited to this.

  Referring to FIG. 12, input pattern information compressor 317 obtains information on the number of pages stored in counter 302, information stored in input pattern information buffer 310, and information stored in representative pattern label buffer 312. The data is mixed and compressed, and supplied to the data mixer 319 (S415). Note that the input pattern information compressor 317 creates input pattern information for each page and performs compression for each page.

  For example, the number from the number 2108 of input patterns included in the 0th page (represented by PC [0]) to the representative pattern number 2109 corresponding to the (PC [0] -1) th input pattern on the 0th page is 1. One compression unit. The number of bytes after compression is stored as the capacity 2106 after compression of the input pattern data of the 0th page. A similar procedure is performed for each page. This is because compression for each page enables decoding for each page, reduces the memory capacity required for decoding, and enables random access of pages.

  Here, the compressed capacity of the input pattern data is stored for each page, and the input pattern information is accessed based on the capacity. Apart from this capacity, an offset to the input pattern information (offset from the number of pages 2105) may be stored for each page, and the input pattern information may be accessed.

  The input pattern information compressor 317 outputs the information from the number of pages 2105 (denoted as P) to the capacity 2107 after compression of the input pattern data of the (P-1) th page without being compressed.

  The input pattern information compressor 317 performs data compression using an entropy coding technique using arithmetic coding. The data compression method is not limited to this.

  The data mixer 319 concatenates the representative pattern information, the representative pattern image, and the compressed data of the input pattern information obtained in S412, S413, and S415, respectively, into one encoded data and outputs the encoded data to the encoded data buffer 320. (S416). Through the above processing, the encoded data buffer 320 stores data obtained by encoding the document image.

  Referring to FIG. 13, the data stored in encoded data buffer 320 includes representative pattern information 2101, representative pattern image 2102, and input pattern information 2103. When the representative pattern information 2101, the representative pattern image 2102 and the input pattern information 2103 are respectively decoded, data as shown in FIGS. 14, 15 and 12 is obtained.

  With reference to FIG. 16, the process of S407 of FIG. 7 will be described in detail. When a character is used as an input pattern as in the present embodiment, the input pattern extractor 305 uses each character instead of outputting a character code corresponding to each character extracted from the paper image in the character recognition technology. Are stored in the input pattern image buffer 309, and the vertical and horizontal widths of the input pattern are stored in the input pattern information buffer 310.

  The character element extraction unit 701 extracts a character element from the image stored in the image data buffer 304, that is, the image of the current page to be processed, and stores information about the circumscribed rectangle of the character element in the character element buffer 702. (S801). The character element indicates a connected component of black pixels (character area pixels). The character element buffer 702 stores the x and y coordinates of the upper left vertex of the circumscribed rectangle and the x and y coordinates of the lower right vertex. An example of such a method for extracting a circumscribed rectangle of a connected component of character area pixels from an image is disclosed in Japanese Patent Laid-Open No. 5-81474. When the processing of S801 is performed, the image is binarized in advance using the binarization threshold stored in the binarization threshold buffer 308.

  The character string direction determination unit 703 refers to the character element buffer 702 to determine whether the direction of the character string in the image is the vertical direction or the horizontal direction, and the determination result is used as the character string direction information flag 713. (S802). An example of a method for determining the direction of a character string in an image from the arrangement of character elements is disclosed in Japanese Patent Laid-Open No. 11-73475.

  The character string extraction unit 705 extracts a character string while referring to the character element buffer 702 and the character string direction information flag 713, and rewrites the contents of the character string information buffer (S803). An example of a method for extracting a character string from the arrangement of character elements is disclosed in the above-mentioned Japanese Patent Application Laid-Open No. 5-81474. The character string information buffer 706 stores character string numbers and character elements in a one-to-one correspondence and is stored as an integer array.

  The character matching unit 704 initializes the character string counter 709 to 0 (S804), and initializes the character counter 710 to 0 (S805). Hereinafter, the value of the character string counter 709 is represented by i, and the value of the character counter 710 is represented by j. The following processing is performed for each character string. That is, the individual character extraction unit 707 divides the i-th character string into character candidate regions (S806). Specifically, it is processed as follows. For example, consider a case where a horizontally written character string “river” is processed. Referring to FIG. 17A, for the character string “river”, the individual character extraction unit 707 divides the character string into individual character candidate regions as indicated by the dotted lines in FIG. This is performed by integrating character elements (in this case, connected components) having circumscribed rectangles overlapping in a direction perpendicular to the character string direction as one character element.

  For example, in FIG. 17A, the circumscribed rectangles of the three points that constitute the “river” overlap each other in the direction perpendicular to the character string direction (in this case, the vertical direction). Therefore, the three circumscribed rectangles are integrated by the individual character extraction unit 707, and the coordinates of the circumscribed rectangle after integration are stored in the individual character information buffer 708 in the same format as the character element buffer 702. When the character string direction information flag 713 indicates horizontal writing, the character candidate area is stored in the order from the left to the right of the character string. In order. The individual character information buffer 708 also stores information on how many characters are present for each character string.

  The character matching unit 704 initializes the character counter 711 in the character string to 0 (S807). Hereinafter, the value of the character counter in character string 711 is represented by k. The character matching unit 704 matches the kth character in the character string with all the character standard patterns 712 while referring to the individual character information buffer 708 and the image data buffer 304, and uses the highest similarity as a matching score. (S808).

  Note that the similarity between the character standard pattern 712 corresponding to each recognition category and the input pattern is calculated based on the composite similarity. For this reason, the maximum value of the similarity is 1. The compound similarity is described in detail in, for example, “Hashimoto“ Introduction to Character Recognition ”(Telecommunications Association, 1982), P35”. A mesh feature amount is used as an example of a feature amount used for calculating the similarity, but it goes without saying that other feature amounts may be used.

  If the matching score is equal to or greater than the predetermined threshold (NO in S809), it is determined that the input pattern has been successfully extracted, and the character matching unit 704 uses the coordinate information of the kth character element as the input pattern information buffer. It is stored in 310 (S812). Also, the character matching unit 704 cuts out a character image from the image data buffer 304 based on the circumscribed rectangle of the character element, and stores it in the input pattern image buffer 309 (S813).

  The character matching unit 704 increments the character counter 711 in the character string by 1 (S814), and if the character string reaches the end of the character string (YES in S815), the character counter 711 is set to the value of the character counter 710. Are added (S816). At this time, the value of the in-character string character counter 711 indicates how many characters have been extracted from the i-th character string that has just been processed.

  The character string counter 709 is incremented by 1 (S817). If the value of the character string counter 709 is different from the number of character strings (NO in S818), since there is an unprocessed character string, the control returns to S806.

  If the value of character string counter 709 is equal to the number of character strings (YES in S818), the processing has been completed for all character strings. For this reason, the value of the character counter 710 is written into the input pattern information buffer 310, and the process ends (S819).

  If the matching score falls below a predetermined threshold (YES in S809), the reintegration / matching process described below is performed (S810). As a result of the reintegration / matching process, the contents of the individual character information buffer 708 are rewritten (S811), and the processes after S812 described above are executed.

  With reference to FIG. 17, S810 (reintegration / matching processing) in FIG. 16 will be described. FIG. 17A shows a character candidate area extracted by the individual character extraction unit 707 from the horizontally written character string “river”. It can be seen that the character area candidates are surrounded by broken lines and five character candidate areas are extracted. Referring to FIG. 17B and FIG. 17F, in the processing of S808, the character matching unit 704 performs matching between the “kawa” and the katakana “shi”. The score is required to be 0.8. A value of 0.8 is not necessarily high. This is because there is a considerable difference in detail between Suzu and Katakana “shi”.

  It is assumed that the threshold used in S809 is 0.85. In this case, since the condition of S809 is not satisfied (YES in S809), the reintegration / matching process (S810) is executed. That is, the character candidate areas are sequentially integrated within a certain character width range. Each time integration is performed, the character matching unit 704 calculates the similarity between all the character standard patterns 712 and the character candidate regions, and extracts the character candidate region having the largest matching score. FIG. 17B, FIG. 17C, and FIG. 17D are character candidate areas each having a certain character width or less, and refer to FIG. 17F, FIG. 17G, and FIG. Thus, the matching scores are determined as 0.5, 0.9, and 0.7, respectively. Among the three character candidate regions, the character candidate region shown in FIG. 17C shows the highest matching score. For this reason, the character candidate area shown in FIG. 17C is adopted as the input pattern.

  By the reintegration / matching process described above (S810 in FIG. 16), the number of characters in the character string of interest decreases. Therefore, the number of characters in the character string and the character coordinates stored in the individual character information buffer 708 are also changed accordingly (S810). For example, in the example used here, the number of characters in the character string is reduced by one by combining the letters “Yes” and extracting the characters “river”. In addition, the “OK” coordinates stored in the individual character information buffer 708 are deleted, and the coordinates corresponding to each other are replaced with the “river” coordinates (S811).

  Next, with reference to FIG. 18, the process of extracting a representative pattern (S410 in FIG. 7) will be described in detail.

  The controller 1000 initializes the representative pattern label buffer 312 (S1101). The representative pattern label buffer 312 is an integer array for associating a representative pattern with an input pattern on a one-to-one basis. The subscript of the representative pattern label buffer 312 corresponds to the input pattern number, and the element of the representative pattern label buffer 312 corresponds to the representative pattern number. Initializing the representative pattern label buffer 312 means substituting different values for each element. Hereinafter, each element of the representative pattern label buffer 312 is represented as LB [i] (i = 0, 1,...), And initialization is performed so that LB [i] = i.

  The loop detector 1001 refers to the binarization threshold buffer 308, the input pattern image buffer 309, and the input pattern information buffer 310, detects the number of loops included in the image corresponding to the input pattern, and loop number buffer 1002 (S1102). A loop is a ring-shaped part. The loop number buffer 1002 is an integer array in which the input pattern number is a subscript and the number of loops is an element. In the following description, the i-th element of the loop number buffer 1002 is L [i]. That is, the number of loops included in the i-th input pattern is represented by L [i].

  In this detection of the number of loops, the pixel of interest is selected based on the same criteria as the input pattern extractor 305. That is, the non-background pixel is set as the target pixel. FIG. 19A shows an image with two loops, and FIG. 19B shows an image with one loop. The process of S1102 will be described in detail later.

  The controller 1000 initializes the first counter 1003 to 0 (S1103), and substitutes a value obtained by adding 1 to the value of the first counter 1003 to the second counter 1004 (S1104). In the following description, the value of the first counter 1003 is i, and the value of the second counter 1004 is j.

  The controller 1000 determines whether or not the i-th and j-th input patterns are similar in size (S1105). This is executed by extracting and comparing the horizontal width and vertical width of the two input patterns.

If the coordinates of the upper left vertex of the circumscribed rectangle of the i th input pattern are (sx0 [i], sy0 [i]) and the coordinates of the lower right vertex are (ex0 [i], ey0 [i]), the i th input The horizontal width lx [i] and the vertical width ly [i] of the pattern, and the horizontal width lx [j] and the vertical width ly [j] of the jth input pattern are represented by the following equations (2) to (5), respectively. .
lx [i] = ex0 [i] −sx0 [i] +1 (2)
ly [i] = ey0 [i] -sy0 [i] +1 (3)
lx [j] = ex0 [j] −sx0 [j] +1 (4)
ly [j] = ey0 [j] −sy0 [j] +1 (5)
At this time, if both the following expressions (6) and (7) hold, it is determined that the i-th and j-th input patterns are similar in size. That is, when the difference between the horizontal width and the vertical width is not so large as compared with the horizontal width or the vertical width itself, it is determined that the sizes are similar.
abs (lx [i] −lx [j]) × 4 ≦ max (lx [i], lx [j]) (6)
abs (ly [i] −ly [j]) × 4 ≦ max (ly [i], ly [j]) (7)
Here, abs (x) indicates the absolute value of x, and max (x, y) indicates the absolute value of x and y.

  If the sizes of the input patterns are similar (YES in S1105), the controller 1000 determines that the number of loops L [i] included in the i-th input pattern and the number of loops L [j] included in the j-th input pattern. Are equal to each other (S1106).

  If the loop numbers L [i] and L [j] are equal (YES in S1106), the pattern comparator 1005 compares the i-th input pattern with the j-th input pattern (S1107).

  If the i-th input pattern is similar to the j-th input pattern (YES in S1105), the controller 1000 rewrites the representative pattern label buffer 312 as follows (S1109). That is, the controller 1000 uses a value min (LB [common to elements LB [i] and LB [j] of the representative pattern label buffer 312 corresponding to the i-th and j-th input patterns determined to be similar to each other. i], LB [j]). A common value min (LB [i], LB [j]) is also substituted for an element having the same value as the element LB [i] or LB [j] before update. Here, min (LB [i], LB [j]) indicates the minimum value of LB [i] and LB [j].

  The controller 1000 increments the second counter 1004 by 1 (S1110). The controller 1000 checks whether or not the value j of the second counter 1004 is equal to the number of input patterns (S1111). If not equal to the number of input patterns (NO in S1111), the controller 1000 returns to S1105.

  If the value j of the second counter 1004 is equal to the number of input patterns (YES in S1111), since all comparisons for the i-th input pattern have been completed, the controller 1000 increments the first counter 1003 by one (S1112). ).

  The controller 1000 checks whether or not the value i of the first counter 1003 is equal to the number of input patterns (S1113). If not equal to the number of input patterns (NO in S1113), the controller 1000 starts comparison with the i-th input pattern. Return to S1004.

  If the value i of the first counter 1003 is equal to the number of input patterns (YES in S1113), the comparison for all combinations of input patterns is complete, and the controller 1000 again sets the first counter 1003 and the second counter 1004. It is initialized to 0 (S1114, S1115).

  The controller 1000 determines whether LB [i] = i is satisfied (S1116). If LB [i] = i (YES in S1116), the controller 1000 reads the image of the i-th input pattern from the input pattern image buffer 309 in order to use the i-th input pattern as the representative pattern. Writing to the buffer 313 is performed (S1117). Further, the controller 1000 reads information on the i-th input pattern from the input pattern information buffer 310 and writes it in the representative pattern information buffer 314 (S1118). In addition, the controller 1000 increments the second counter 1004 by 1 (S1119).

  The reason why the i-th input pattern is used as a representative pattern only when the condition LB [i] = i is satisfied is to select only one representative pattern from the input patterns belonging to the same cluster. In addition, if there is a method for selecting only one representative pattern from the input patterns, that method may be used.

  In the processing of S1118, the horizontal width lx [i] and the vertical width ly [i] of the i-th input pattern are obtained according to the above equations (2) and (3).

  The controller 1000 increments the first counter 1003 by 1 (S1120). The controller 1000 checks whether or not the value i of the first counter 1003 matches the number of input patterns (S1121). If they do not match (NO in S1121), the process returns to S1116.

  If the two match (YES in S1121), with reference to FIG. 14, the controller 1000 writes the value j of the second counter 1004 in the representative pattern information buffer 314 as the number of representative patterns 2104 (S1122). The process of repacking the value of the representative pattern label buffer 312 is performed (S1123).

  Referring to FIG. 15, representative pattern image buffer 313 has image data of a representative pattern written in the order of raster scan. Such a data structure is merely an example, and it goes without saying that other data structures may be used.

  The process of S1123 will be described. There are j representative patterns. However, the elements of the representative pattern label buffer 312 can take values ranging from 0 to “number of input patterns−1”. For this reason, the element of the representative pattern label buffer 312 takes a jump value. The controller 1000 replaces the elements so that the elements of the representative pattern label buffer 312 are within the range from 0 to j−1 and the magnitude relationship of the elements is maintained. For example, referring to FIG. 20A, when the value of the representative pattern label buffer 312 having elements 0, 2 and 5 is repacked, the result is as shown in FIG.

  The representative pattern may be created from a plurality of input patterns belonging to the same cluster. For example, the input pattern may be enlarged or reduced to make the size equal, and the average of the input pattern may be taken to create a representative pattern. However, in general, such synthesis processing often results in an image with a blurred representative pattern, and is not always effective.

  An example of a change in the value of the representative pattern label buffer 312 will be described with reference to FIGS. FIG. 21 is a diagram in which two types, a total of 13 input patterns, are arranged on the pattern space. The numbers in the figure indicate the values of the representative pattern labels stored in the representative pattern label buffer 312 immediately after S1102. This number also matches the input pattern number stored in the input pattern information buffer 310.

  FIG. 22 to FIG. 34 show the change in the value of the representative pattern label after the processing from S1104 to S1112 is executed while the value i of the first counter 1003 is incremented by one. The process of S1107 is executed for all combinations of input patterns. Here, it is assumed that whether or not two input patterns are similar is determined by whether or not the Euclidean distance in the pattern space is equal to or less than a certain threshold value, and the circle indicated by the dotted line is the pattern located at the center. The range of distance determined to be similar to is shown. For example, FIG. 22 shows the state of the representative pattern label buffer 312 when the process of S1112 is completed when i = 0. Since it is determined that the 0th input pattern 2801 and the first input pattern 2802 are similar, the value of the representative pattern label buffer 312 is rewritten as shown in FIG. 22 by the processing of S1109.

  In the initial state shown in FIG. 21, LB [0] = 0 and LB [1] = 1, but LB [1] is rewritten to the same value as LB [0] by the processing of S1109. There is no other representative pattern label with the same value as LB [0] or LB [1] immediately before S1109 is executed. For this reason, other representative pattern labels maintain the values as they are. For example, since the ninth input pattern 2803 is not determined to be similar to the input pattern 2801, the representative pattern label LB [9] remains 9 at this time. 23 to 34 show values of the representative pattern label buffer 312 when the process of S1112 is completed when i changes from 1 to 12. FIG. For example, in FIG. 33 corresponding to i = 11, it is determined that the 11th input pattern 2804 and the 12th input pattern 2805 are similar. For this reason, LB [11] = 0 and LB [12] = 0 at the end of the processing of S1112. Further, as shown in the figure, it is not determined that the seventh input pattern 2806 is similar to the input pattern 2804, but LB [7] = LB [11] in the processing so far. It is shown. For this reason, LB [7] is also rewritten to 0 by the processing of S1109.

  FIG. 34 is a diagram corresponding to i = 12, but LB [12] has already been rewritten to 0, and the representative corresponding to the input pattern determined to be similar to the 12th representative pattern 2807. The values of the pattern labels are all 0. For this reason, the value of the representative pattern label buffer 312 does not change.

  The state at the time when the process of S1114 is executed is shown in FIG. FIG. 36 shows the state after the processing of S1123. By executing the refill process (S1123) of the representative pattern label buffer 312, the value 3 of the representative pattern label is updated to 1.

  With reference to FIG. 37, the process of S1102 of FIG. 18 will be described in detail. The controller 1300 initializes the first counter 1301 to 0 (S1401). Hereinafter, the value of the first counter 1301 is i. The value of the first counter 1301 indicates the number of the input pattern currently being processed by the loop detector 1001.

The controller 1300 uses the connected component circumscribed rectangle extractor 1302 to extract the connected component of the background area from the input pattern indicated by the first counter 1301. The controller 1300 creates a circumscribed rectangle for each connected component and stores the information in the connected component circumscribed rectangle information buffer 1303 (S1402). That is, the page number p [i] to which the i-th input pattern belongs is extracted from the input pattern information buffer 310. The pixel of interest is a pixel that satisfies the following expression (8), where TH [p [i]] is the binarization threshold value of the image of the p [i] page stored in the binarization threshold buffer 308. This can be done. This means that the pixel of interest is not a non-background area but a background area. In other respects, the operation of the connected component circumscribed rectangle extractor 1302 may be the same as that disclosed in JP-A-5-81474.
TH [p [i]] ≦ pixel value <256 (8)
The connected component circumscribed rectangle information buffer 1303 stores the number of rectangles RC, the X and Y coordinates of the upper left vertex of each rectangle, and the X and Y coordinates of the lower right vertex of each rectangle. Hereinafter, the upper left vertex of the kth rectangle is represented as (sx1 [k], sy1 [k]), and the lower right vertex is represented as (ex1 [k], ey1 [k]).

The controller 1300 initializes the i-th element L [i] of the loop number buffer 1002 to 0 (S1403), and initializes the second counter 1304 to 0 (S1404). Here, the value of the second counter 1304 is represented as j. The controller 1300 checks whether or not the jth rectangle included in the connected component circumscribed rectangle information buffer 1303 is in contact with the end of the input pattern (S1405 to S1408). That is, the XY coordinates of the upper left vertex and lower right vertex of the circumscribed rectangle of the i-th input pattern stored in the input pattern information buffer 310 are (sx0 [i], sy0 [i]) and (ex0 [i], ey0 [i]), it is checked whether any one of the equations (9) to (12) holds.
sx1 [j] = 0 (9)
sy1 [j] = 0 (10)
ex1 [j] = ex0 [i] -sx0 [i] (11)
ey1 [j] = ey0 [i] −sy0 [i] (12)
If any of the four conditions is satisfied (YES in S1405, YES in S1406, YES in S1407, or YES in S1408), controller 1300 increments second counter 1304 by one (S1410).

  If neither condition is satisfied (NO in S1405, NO in S1406, NO in S1407, and NO in S1408), the controller 1300 increments the i-th element L [i] of the loop number buffer 1002 by one. (S1409), the process proceeds to S1410.

  After executing the processing of S1410, the controller 1300 checks whether or not the value j of the second counter 1304 matches the number of rectangles extracted by the connected component circumscribed rectangle extractor 1302 (S1411). If they match (YES in S1411), the controller 1300 increments the first counter 1301 by 1 (S1412). If they do not match (NO in S1411), the process returns to S1405.

  After executing the processing of S1412, the controller 1300 checks whether or not the value i of the first counter 1301 matches the number of input patterns (S1413). If the value i of the first counter 1301 matches the number of input patterns (YES in S1413), the process ends. If they do not match (NO in S1413), the process returns to S1402.

  An example of processing of the loop detector 1001 will be described with reference to FIG. When FIG. 38A is used as the input pattern, FIG. 38B is a diagram in which the non-background region and the background region are reversed and displayed. In FIG. 38B, the background region is represented in black, and is the region of interest of the connected component circumscribed rectangle extractor 1302. FIG. 38C shows connected components 1501 and 1502 whose circumscribed rectangle is in contact with the end of the input pattern among the connected components extracted from the image of FIG. FIG. 38D shows connected components 1503 and 1504 whose circumscribed rectangle is not in contact with the end of the input pattern among the connected components extracted from the image of FIG.

  By counting the number of connected components whose circumscribed rectangles, such as connected components 1503 and 1504, do not touch the edge of the input pattern, the number of loops in the non-background region can be calculated.

  In this way, the calculation of the number of loops in the non-background region focuses on the background region and counts only the connected component circumscribed rectangles that have not reached the end. This is easier than the method of counting the number of detected loops after the detection.

  In addition, by adopting such a configuration, conditions are imposed on the size and shape of the opening of the loop to be detected. For example, adding a process of “ignoring the number of loops having a horizontal width or a vertical width equal to or less than a certain value” also calculates ex1 [j] −sx1 [j], ey1 [j] −sy1 [j] This can be done easily by ignoring the ones that do not meet the conditions. Any other conditions are the same as long as they can be replaced by conditions related to the size and shape of the circumscribed rectangle of the opening of the loop.

  With reference to FIG. 39, the process of S1107 of FIG. 18 will be described. The vector converter 1601 performs feature extraction from each of the two input patterns to be compared, and generates a feature vector (S1701). Various methods for feature extraction have been proposed in the field of character recognition. Here, as an example, feature extraction is performed by the method described below, and the input pattern is converted into a feature vector.

  Referring to FIG. 40A, an input pattern of 3 × 5 pixels is divided into four equal parts. The pixels included in each section are weighted by the area ratio included in the section, and a weighted average is obtained. The value of the weighted average is as shown in FIG. 40B, and a four-dimensional feature vector is created from this. In practice, a 64-dimensional (8 × 8-dimensional) feature vector as shown in FIG. 41A is calculated.

  The vector normalizer 1602 normalizes the feature vector so that the absolute value becomes 1 (S1702). That is, the vector normalizer 1602 obtains the absolute value of the feature vector and divides each element of the feature vector by the absolute value.

The vector canonicalizer 1603 normalizes the feature vector (S1703). Here, the canonicalization means that a feature vector having the same elements, an absolute value of 1 is C, and a feature vector of an input pattern created in the process of S1701 is F. From these, The feature vector F ′ is calculated based on the above.
F ′ = F− (C · F) C (13)
C · F represents the inner product of feature vectors. The feature vector F ′ is obtained by extracting a component orthogonal to the input pattern having a uniform background. The reason for canonicalization is as follows. In a document image or the like, since black characters are written on a white background, the background pixel value indicates a large value. In particular, in the case of simple characters, most of the image shows a large value, and the feature vector is similar to that created from an input pattern of uniform density regardless of the type of input pattern. . Canonicalization is performed to prevent this. The general theoretical basis of canonicalization is described in detail in, for example, Iijima “Pattern Recognition Theory” (Moritakita Publishing, 1989), P94.

  The inner product calculator 1604 calculates the inner product S0 of the two feature vectors obtained in the processing of S1703 (S1704). The inner product here is a value obtained by dividing the sum of the products of the corresponding elements by the product of the absolute values of the two feature vectors, and takes a value in the range of 0 to 1. As the inner product value S0 is closer to 1, the two feature vectors are more similar, indicating that the two input patterns are similar.

  The controller 1600 checks whether or not the inner product value S0 is equal to or greater than a predetermined threshold value TH0 (S1705). If inner product value S0 is smaller than threshold value TH0 (NO in S1705), it is determined that they are not similar (S1710), and the process is terminated.

  If the inner product value S0 is equal to or greater than the threshold value TH0 (YES in S1705), the controller 1600 sets the counter 1605 to 0 in order to compare the feature vector portions (hereinafter referred to as “partial vectors”). Initialization is performed (S1705). Hereinafter, the value of the counter 1605 is k.

  A partial vector refers to a vector created by extracting some of the elements of a feature vector. Here, nine 16-dimensional vectors as shown in FIGS. 41B to 41J are used as partial vectors from 64-dimensional feature vectors as shown in FIG. Numbers from 0 to 8 are assigned to the partial vectors in FIGS. 41 (b) to (j), respectively.

  The partial vector generator 1606 generates a k-th partial vector for each of the two feature vectors (S1706). The inner product calculator 1604 calculates the inner product S1 [k] between the partial vectors (S1707). The controller 1600 checks whether or not the inner product S1 [k] is greater than or equal to a predetermined threshold value TH1 (S1709). If inner product S1 [k] is smaller than threshold value TH1 (NO in S1709), it is determined that they are not similar (S1708), and the process is terminated.

  If inner product S1 [k] is equal to or greater than threshold value TH1 (YES in S1709), controller 1600 increments counter 1605 by one (S1709). If the value k of the counter 1605 does not match the number of partial vectors (NO in S1712), the process returns to S1707.

  If the value k of the counter 1605 matches the number of partial vectors (YES in S1712), since it is determined that all partial vectors are similar, it is determined that the two input patterns are similar (S1713). ), The process is terminated.

  Note that the threshold value TH0 and the threshold value TH1 can be determined independently. It is also possible to set different threshold values for every nine partial vectors. Empirically, better results are often obtained when the threshold value TH0 is larger than the threshold value TH1. This is because, in the comparison between a plurality of partial vectors, if the degree of similarity between all the partial vectors is not equal to or greater than a certain value, it is not determined that the vectors are similar, and the comparison of the partial vectors is strict. . As an example, threshold values TH0 and TH1 are set to 0.9 and 0.8, respectively.

  Here, as an indicator of whether two patterns are similar, an inner product indicating that the two patterns are more similar as the value is higher is used, but a scale indicating that the two patterns are more similar as the value is smaller. Alternatively, the Euclidean distance or the city block distance between the partial vectors may be used. The same applies to the comparison of feature vectors in S1705.

  The reason for comparing the parts is as follows. That is, it is for correctly identifying patterns that are similar on the whole but different on a partial basis. FIG. 42 (a) and FIG. 42 (b) show an example of such a similar pattern. Even in such a pattern, it can be seen that when only the upper right portion is extracted as shown in FIGS. Therefore, by requesting that two patterns are similar for all partial vectors, correct identification can be performed, and patterns representing different characters as shown in FIGS. It is possible to prevent replacement with a common representative pattern.

  The reason why loop detection is performed when extracting a representative pattern is to correctly identify an input pattern that is difficult to identify even if partial comparison is performed. For example, in the example as shown in FIGS. 43A and 43B, unlike the case shown in FIG. 42, not only is it similar as a whole, but also in the upper right part where the difference is considered to be the largest, Similar to FIGS. 43 (c) and 43 (d). However, even in such a case, the number of loops is different. Therefore, it is possible to prevent the patterns representing different characters from being replaced with a common representative pattern as shown in FIGS. 43 (a) and 43 (b).

  With reference to FIG. 44, a decoding process of encoded data will be described. The data separator 2202 separates the encoded data stored in the encoded data buffer 2201 into representative pattern information 2101, representative pattern image 2102, and input pattern information 2103 shown in FIG. The data separator 2202 transmits the separated representative pattern information 2101, representative pattern image 2102 and input pattern information 2103 to the representative pattern information decompressor 2203, representative pattern image decompressor 2204 and input pattern compression information buffer 2205, respectively (S2301). .

  The representative pattern information decompressor 2203 decompresses the representative pattern information 2101 and stores it in the representative pattern information buffer 2206. (S2302). The representative pattern image expander 2204 expands the representative pattern image 2102 and stores it in the representative pattern image buffer 2207 (S2303). At this time, the representative pattern information buffer 2206 stores data as shown in FIG. 14, and the representative pattern image buffer 2207 stores data as shown in FIG.

  The representative pattern pixel value converter 2208 restores the pixel value of the representative pattern stored in the representative pattern image buffer 2207 using the pixel value conversion table 2209 (S2304). This is a process of returning the pixel value reduced during encoding to the pixel value of the number of gradations before encoding. FIG. 45 shows an example of the pixel value conversion table 2209. The first row shows the input pixel value, and the second row shows the corresponding output pixel value.

  Based on the data stored in the representative pattern information buffer 2206, the representative pattern image offset generator 2210 sets the storage position of each representative pattern in the representative pattern image buffer 2207 as an offset from the head of the representative pattern image buffer 2207. calculate. The representative pattern image offset generator 2210 stores the offset in the representative pattern image offset table 2211 which is an integer array in which the representative pattern number and the offset value have a one-to-one correspondence (S2305). The product of the horizontal width and the vertical width of each representative pattern stored in the representative pattern information buffer 2206 directly indicates the capacity after expansion of each representative pattern. For this reason, the offset can be easily calculated.

  Referring to FIG. 12, input pattern information offset generator 2212 calculates the number of pages 2105 (represented as P) at the head of input pattern compression information buffer 2205 after compression of input pattern data on page P−1. By referring to the capacity 2107, it is calculated where in the input pattern compression information buffer 2205 each page data starts. The input pattern information offset generator 2212 writes the calculation result in the input pattern information offset table 2213, which is an integer array in which page numbers and page data storage locations have a one-to-one correspondence (S2306). For example, the offset of the input pattern data corresponding to the i-th page is obtained as the sum of the compressed capacities of the input pattern data from the 0th page to the (i−1) -th page.

  The page counter 2214 is initialized to 0 (S2307). Hereinafter, the value of the page counter 2214 is i. The page image buffer initializer 2216 initializes the pixel value of the image stored in the page image buffer 2215 to the same value as the background color (S2308). Here, the value of the background color is represented by 255. Here, the background color of the image stored in the page image buffer 2215 is a fixed value, but the background color may also be encoded to change the background pixel value.

  The input pattern information decompressor 2217 refers to the input pattern compression information buffer 2205 and the input pattern information offset table 2213, decompresses the input pattern information included in the i-th page, and stores it in the input pattern information buffer 2218 (S2309). ). The input pattern counter 2219 is initialized to 0 (S2310). Hereinafter, the value of the input pattern counter 2219 is represented by j.

  The pixel density converter 2220 calculates the horizontal width and vertical width of the j-th input pattern of the i-th page from the data stored in the input pattern information buffer 2218 (S2311).

  From the data stored in the input pattern information buffer 2218 and the representative pattern information buffer 2206, the horizontal width and vertical width of the input pattern and the representative pattern representing the input pattern are extracted, and the horizontal width and vertical width are respectively compared. If either or both of the horizontal width and the vertical width do not match (NO in S2312), the pixel density converter 2220 makes the non-matching one or both of the vertical width and horizontal width of the representative pattern match that of the input pattern. (S2313). As a method for converting the image size, a method such as bilinear interpolation has been conventionally proposed. Since these methods are known techniques, detailed description thereof will not be repeated here. For such a method, Takagi et al., “Image Analysis Handbook” (The University of Tokyo Press, 1991) PP. 441-PP. 444 in detail.

  After the horizontal width and vertical width of the representative pattern are matched with the input pattern (YES in S2313 or S2312), the representative pattern is inserted into the position where the input pattern of the page image buffer 2215 exists (S2314).

  Here, the processing of S2313 is omitted only when the sizes completely match, but the conditions are further relaxed, and when the difference between the horizontal width and the vertical width is small, omitting S2313 has a significant effect on the image quality. Can be speeded up without giving

  The input pattern counter 2219 is incremented by 1 (S2114). It is checked whether or not the value j of the input pattern counter 2219 matches the number of input patterns on the i-th page (S2316). If the value j of the input pattern counter 2219 does not match the number of input patterns (YES in S2316), the same processing is repeated for the remaining input patterns, and the control returns to S2311.

  If the value j of the input pattern counter 2219 matches the number of input patterns (YES in S2316), the process for the i-th page is completed, and the page counter 2214 is incremented by 1 (S2319).

  It is checked whether or not the value i of the page counter 2214 matches the number of pages (S2318). If the number i does not match the number of pages (NO in S2318), the process is performed on the remaining pages, so the control is S2308. Return to.

  If it matches the number of pages (YES in S2318), the processing is completed for all pages, so after outputting the image (S2319), the processing is terminated.

  As described above, according to the embodiment of the present invention, an input pattern is represented as a feature vector, and partial vectors constituting the feature vector are compared with each other. Thus, by partially comparing the input patterns, it is possible to distinguish characters that are similar in overall view but are not similar in partial view. For this reason, the replacement error of the input pattern can be reduced.

  Also, by detecting the number of loops, although partially similar, it is possible to accurately distinguish different characters. For this reason, the replacement error of the input pattern can be reduced.

  Furthermore, the number of representative patterns representing an input pattern can be reduced by chaining the similarity range of the input pattern. For this reason, encoding efficiency can be kept high.

  Furthermore, characters cut out from the image data are used as the representative pattern. Therefore, the input pattern replacement error due to character recognition does not occur as in the case where the input pattern is recognized and the representative pattern is represented by the character code.

  Further, unlike the case where the connected component is used as an input pattern, there is no sense of incongruity in the decoded image.

  At the time of image decoding, an image can be created by simply pasting the representative pattern to the coordinate position of the input pattern. For this reason, an image can be restored at high speed.

  Further, since the encoding unit of the coordinate position of the input pattern corresponds to the page of the document, only the image corresponding to the desired page can be easily decoded.

  In the calculation method of the number of loops included in a figure according to the present invention, the calculation of the number of loops in the non-background region is a circumscribed rectangle of the connected component detected from the background region and does not reach the end. By counting only, it is easier to perform than the conventional method of actually detecting a loop and counting the number of detected loops.

  In addition, by adopting such a configuration, it is easy to place conditions on the shape and size of the opening of the loop to be detected as long as it can be replaced with the conditions on the shape and size of the circumscribed rectangle of the loop opening. Can be realized.

  The above-described image encoding device and image decoding device can be realized by a computer and a program operating on the computer. The image encoding processing program and the image decoding processing program may be provided by a computer-readable recording medium such as a CD-ROM (Compact Disc-Read Only Memory), and the computer may read and execute the program. . The computer may receive a program distributed via a network and execute the received program.

  The embodiment disclosed this time should be considered as illustrative in all points and not restrictive. The scope of the present invention is defined by the terms of the claims, rather than the description above, and is intended to include any modifications within the scope and meaning equivalent to the terms of the claims.

It is a figure which shows the structure of the image coding apparatus which concerns on embodiment of this invention. 3 is a block diagram showing a configuration of an input pattern extractor 305. FIG. 5 is a block diagram showing a configuration of a representative pattern extractor 311. FIG. 2 is a block diagram showing a configuration of a loop detector 1001. FIG. 4 is a block diagram showing a configuration of a pattern comparator 1005. FIG. It is a block diagram which shows the structure of the image decoding apparatus which concerns on embodiment of this invention. It is a flowchart of an image encoding process. 4 is a diagram illustrating an example of data stored in an image data buffer 304. FIG. It is the figure which expanded a part of input image. It is a figure which shows the input pattern obtained from the input image. It is a figure which shows the character cut out from the input image. 5 is a diagram illustrating an example of input pattern information 2103. FIG. 4 is a diagram illustrating an example of data stored in an encoded data buffer 320. FIG. 5 is a diagram illustrating an example of representative pattern information 2101. FIG. 5 is a diagram illustrating an example of a representative pattern image 2102. FIG. It is a flowchart of the process which extracts an input pattern from a binary image. It is a figure for demonstrating the specific example of the process which extracts an input pattern from a character string. It is a flowchart of the process which extracts a representative pattern. It is a figure which shows an example of the input pattern from which the number of loops differs. FIG. 10 is a diagram for explaining a process of repacking values in a representative pattern label buffer 312; It is a figure for demonstrating the change of the value of the representative pattern label buffer 312. FIG. It is a figure for demonstrating the change of the value of the representative pattern label buffer 312. FIG. It is a figure for demonstrating the change of the value of the representative pattern label buffer 312. FIG. It is a figure for demonstrating the change of the value of the representative pattern label buffer 312. FIG. It is a figure for demonstrating the change of the value of the representative pattern label buffer 312. FIG. It is a figure for demonstrating the change of the value of the representative pattern label buffer 312. FIG. It is a figure for demonstrating the change of the value of the representative pattern label buffer 312. FIG. It is a figure for demonstrating the change of the value of the representative pattern label buffer 312. FIG. It is a figure for demonstrating the change of the value of the representative pattern label buffer 312. FIG. It is a figure for demonstrating the change of the value of the representative pattern label buffer 312. FIG. It is a figure for demonstrating the change of the value of the representative pattern label buffer 312. FIG. It is a figure for demonstrating the change of the value of the representative pattern label buffer 312. FIG. It is a figure for demonstrating the change of the value of the representative pattern label buffer 312. FIG. It is a figure for demonstrating the change of the value of the representative pattern label buffer 312. FIG. It is a figure for demonstrating the change of the value of the representative pattern label buffer 312. FIG. It is a figure for demonstrating the change of the value of the representative pattern label buffer 312. FIG. It is a flowchart of the process which detects the number of loops contained in an input pattern. 6 is a diagram for explaining an example of processing by a loop detector 1001. FIG. 10 is a flowchart of input pattern comparison processing by the pattern comparator 1005; It is a figure for demonstrating the process which extracts the feature-value from an input pattern. It is a figure for demonstrating the relationship between a feature vector and a partial vector. It is a figure for demonstrating a partially different pattern. It is a figure for demonstrating the pattern from which the number of loops differs. It is a flowchart of a decoding process of encoded data. It is a figure which shows an example of the pixel value conversion table 2209. It is a figure for demonstrating an example of the replacement error of an input pattern. It is a figure which shows distribution of an input pattern. It is a figure for demonstrating the encoding of the conventional input pattern. It is a figure for demonstrating the encoding of the conventional input pattern. It is a figure for demonstrating the encoding of the conventional input pattern. It is a figure for demonstrating the encoding of the conventional input pattern. It is a figure for demonstrating the encoding of the conventional input pattern. It is a figure for demonstrating the problem of the encoding of the conventional input pattern.

Explanation of symbols

  301 Auto Feeder, 302 Counter, 303 Scanner, 304 Image Data Buffer, 305 Input Pattern Extractor, 306 Page Counter, 307 Binary Threshold Calculator, 308 Binary Threshold Buffer, 309 Input Pattern Image Buffer, 310 Input Pattern information buffer, 311 representative pattern extractor, 312 representative pattern label buffer, 313 representative pattern image buffer, 314 representative pattern information buffer, 315 representative pattern information compressor, 316 representative pattern image color reducer, 317 input pattern information compressor, 318 Representative pattern image compressor, 319 data mixer, 320 encoded data buffer.

Claims (10)

An input pattern extracting means for extracting from the image data as a pattern that is extracted from the image data and constituting a character or a part constituting the character ;
With the first input pattern extracted by said input pattern extraction means, similar to the third input pattern similar to the second input pattern similar to the first input pattern, to the first input pattern Similar enlargement means for making the input pattern
Representative pattern extraction means for extracting a representative pattern representing a class for classifying the input pattern from among the input patterns determined to be similar to each other by the similarity enlargement means;
An image encoding device comprising: an image of the representative pattern and an output means for outputting a coordinate position of the input pattern.   The image coding apparatus according to claim 1, wherein the representative pattern extraction unit compares the input patterns for each portion constituting the input pattern, and extracts the representative pattern from among similar input patterns. A pattern extracted from the image data and extracting a character or a part constituting the character from the image data as an input pattern;
With the first input pattern extracted, the step of the input pattern the third input pattern similar to the second input pattern similar to the first input pattern, similar to the first input pattern When,
Extracting a representative pattern representing a class for classifying the input pattern from among input patterns similar to each other;
Outputting an image of the representative pattern and a coordinate position of the input pattern.   The image code according to claim 3, wherein the step of extracting the representative pattern includes a step of comparing the input patterns for each portion constituting the input pattern and extracting the representative pattern from among the input patterns similar to each other. Method. An image decoding method for decoding an image from data encoded by the image encoding method according to claim 3 or 4,
Extracting the image of the representative pattern and the coordinate position of the input pattern from the encoded data;
Pasting a representative pattern representing the input pattern to the coordinate position of the input pattern. A computer-readable recording medium that records a computer-executable image encoding method program,
The image encoding method includes:
A pattern extracted from the image data and extracting a character or a part constituting the character from the image data as an input pattern;
With the first input pattern extracted, the step of the input pattern the third input pattern similar to the second input pattern similar to the first input pattern, similar to the first input pattern When,
Extracting a representative pattern representing a class for classifying the input pattern from among input patterns similar to each other;
A computer-readable recording medium comprising: outputting an image of the representative pattern and a coordinate position of the input pattern. The step of extracting the representative pattern includes:
The computer-readable recording medium according to claim 6, comprising comparing the input patterns for each portion constituting the input pattern and extracting a representative pattern from among the input patterns similar to each other. A computer-readable recording medium that records a computer-executable image decoding method program,
The image decoding method includes:
An image decoding method for decoding an image from data encoded by the image encoding method according to claim 3 or 4,
Extracting the image of the representative pattern and the coordinate position of the input pattern from the encoded data;
Pasting a representative pattern representing a class for classifying the input pattern at the coordinate position of the input pattern. An image decoding apparatus that decodes an image from data encoded by the image encoding apparatus according to claim 1 or 2,
An input pattern extracting means for extracting from the image data as an input pattern a pattern extracted from the image data and constituting a character or a character ;
Representative pattern extraction means for comparing the extracted input patterns for each part constituting the input pattern and extracting a representative pattern representing a class for classifying the input pattern from among similar input patterns;
An image decoding apparatus comprising: an output unit that outputs an image of the representative pattern and a coordinate position of the input pattern. An input pattern extracting means for extracting from the image data as a pattern that is extracted from the image data and constituting a character or a part constituting the character ;
With the first input pattern extracted by said input pattern extraction means, similar to the third input pattern similar to the second input pattern similar to the first input pattern, to the first input pattern Similar enlargement means for making the input pattern
Representative pattern extraction means for extracting a representative pattern representing a class for classifying the input pattern from among the input patterns determined to be similar to each other by the similarity enlargement means;
Output means for outputting the image of the representative pattern and the coordinate position of the input pattern,
The image encoding device, wherein the output means outputs an integer array for associating the representative pattern with the input pattern.

Images