JP2013190911A - Character recognition device, recognition dictionary generation device and normalization method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To normalize a plurality of input images with different size so that variation in the shape among the same characters becomes small in order to raise accuracy of character identification.SOLUTION: Preprocessing for reducing disturbance factors from input images is executed, contours of the images to which the preprocessing is executed are extracted, the images to which the preprocessing is executed are synthesized with the images of the extracted contours, mapping from the synthesized image to normalized images with predetermined size are generated so that an image of the center of gravity of the synthesized image approaches the center of a range of the predetermined size, and an image within a range to which pixels of the synthesized image spread approaches the range of the predetermined size, the images to which the preprocessing is executed are normalized according to the generated mapping, the normalized images are converted into vector values on a vector space, the vector values are determined that they are any characters based on a recognition dictionary stored in a storage device, and a result of determination is output.

Description

  The present invention relates to a recognition dictionary generation device and a character recognition device for character recognition, and more particularly to a character image normalization technique.

  The character recognition device determines a character category written in the input image using the recognition dictionary and outputs a determination result. Here, the character category is, for example, 10 numbers from 0 to 9 in the case of number recognition. The recognition dictionary is created by a recognition dictionary generation device.

  The processing flow from when the character recognition device receives an input image to when it outputs a determination result includes four processing procedures: preprocessing, normalization, feature extraction, and identification.

  FIG. 2 is a flowchart showing processing executed by a conventional character recognition device.

  In the character image input unit 201, an image is input by a program executed by a user or an arithmetic device.

  In the preprocessing unit 202, for example, a procedure for removing as much disturbance factors as possible to prevent character recognition is performed by performing noise removal of the input image, smoothing by blurring processing, and the like.

  Next, the normalization unit 203 receives preprocessed images of various sizes as input, and executes a procedure for aligning the sizes of these images. As a result, subsequent processing can be unified.

  Next, the feature extraction unit 209 executes a procedure for receiving the normalized image as an input and converting it into a vector value on a vector space. The vector space is called a feature space, and the vector value is called a feature vector. As a feature extraction method, a method of extracting a pixel feature, a contour feature, a gradient feature, a Gabor feature, etc. is widely known (Non-Patent Document 1). If necessary, the number of dimensions of the feature space is reduced by using a dimension reduction technique such as principal component analysis or linear discriminant analysis (Non-Patent Document 2).

  By the processing so far, the input image is expressed as a vector value (feature vector) on the feature space.

  Next, the identification unit 210 uses the recognition dictionary 214 to execute processing for determining the character category to which the feature vector belongs. The recognition dictionary 214 holds information for determining which character category each point on the feature space belongs to. Details of determination using the recognition dictionary 214 are described in Non-Patent Document 1 or Non-Patent Document 2.

  The output unit 211 outputs the determination result to a display device such as a display or a file.

  In order to perform highly accurate character recognition, the processes of the preprocessing unit 202, normalization unit 203, feature extraction unit 209, and identification unit 210 have an important role. Therefore, it is important to perform processing suitable for character recognition in each processing.

  Characters in the input image to be recognized vary in size, shape, or degree of degradation, even if they are the same character type, depending on the writing instrument, writer, font, scanning environment, paper quality, or paper storage status. is there. In addition to making the sizes of input images uniform, normalization has the purpose of suppressing variations in shape between the same character types in such input images. Thereby, the identification rate of the character recognition device can be increased.

  Existing normalization methods for character images include linear normalization method, non-linear normalization method, moment normalization method, bi-moment normalization method, CBA method, MCBA method, and LDPF method. Among these methods, the moment recognition method and the bi-moment normalization method have published a benchmark result of character recognition that has a high recognition rate (non-patent document 3).

  FIG. 5 is an explanatory diagram of an example of an image generated by the moment normalization method and the bi-moment normalization method. Specifically, in FIG. 5, the bi-moment normalization method is applied to the input image 501 of normalization processing, the normalized image 502 generated by applying the moment normalization method to the input image 501, and the input image 501. A normalized image 503 generated by applying is shown.

  As described above, normalization methods such as the moment method and the bi-moment method are known to have high discrimination ability. However, since these methods calculate moments by directly using the pixel values of the original image, they are easily affected by the thickness of the character stroke. Therefore, the value of the moment varies greatly depending on the thickness of the character, and therefore the position of the character in the normalized image varies depending on the thickness of the character.

  FIG. 6 is an explanatory diagram of an example of different fonts of the same character, and specifically shows an image 601 of the character “T” of a different font. As shown in FIG. 6, the thickness of the character is not essential in determining the character. For this reason, variation in characters (character position, size, etc.) in the normalized image between the same character categories caused by differences in character thickness is disadvantageous for identification.

  The contour feature amount moment normalization method (Patent Document 1, Non-Patent Document 4, Non-Patent Document 5) is a method of extracting the outline of a character and performing normalization based on the moment of the outline of the character. This method is effective for reducing variations between characters due to the length and thickness of characters, and has a high identification rate in type character recognition.

  FIG. 10 is an explanatory diagram of an image normalized by the moment normalization method and the contour feature amount moment normalization method.

  The original image 1001 illustrated in FIG. 10 is a plurality of “T” character images having different horizontal bar thicknesses. Furthermore, an image 1002 obtained by normalizing the original image 1001 by the moment normalization method is shown. In the normalized image 1002, it can be seen that as the horizontal bar becomes thicker, the center of the image is shifted to a portion above T, and the position of the character is lowered downward as a whole. Further, in the normalized image 1002, the thickness of the vertical bar of T, which has the same thickness in the original image 1001, also varies. Such variation appears as variation of vector points on the feature space after feature extraction, and causes a reduction in the identification rate. In contrast, in an image 1003 obtained by normalizing the original image 1001 by the contour feature amount moment normalization method, these variations are reduced.

JP 2010-108113 A

Mohammed Cheriet, Nawwaf Kharma, Cheng lin Liu, and Ching Suen, "Character Recognition Systems: A Guide for Students and Practitioners", Wiley-Interscience, 2007. Kenichiro Ishii, Noriyoshi Ueda, Eisaku Maeda, Hiroshi Murase, “Pattern Recognition”, Ohm Publishing Company, August 1998 Cheng-Lin Liu, Kazuki Nakashima, Hiroshi Sako, and Hiromichi Fujisawa, "Handwritten digit recognition: investigation of normalization and feature extraction techniques", Pattern Recognition, Vol.37, No.2, pp. 265_279, 2004. Toshinori Miyoshi, Takeshi Nagasaki, and Hiroshi Shinjo, "Character Normalization Methods using Moments of Gradient Features and Normalization Cooperated Feature Extraction", Proceedings of the 2009 Chinese Conference on Pattern Recognition and the First CJK Joint Workshop on Pattern Recognition, pp.934-938 , 2009. Toshinobu Miyoshi, Takeshi Nagasaki, Hiroshi Shinjo, “Text Normalization Method Using Gradient Feature Moments”, IEICE Technical Report, PRMU, Pattern Recognition / Media Understanding 108 (432), pp.187-192 , 2009.

  The contour feature amount moment normalization method is a method of extracting a character contour and performing normalization based on the extracted moment value of the character contour portion. This method is effective in suppressing variations in character thickness and length, and is particularly effective in recognition of type characters. However, in some handwritten characters and some printed characters, the outline of the characters may be lost.

  FIG. 13 is an explanatory diagram illustrating an example of a character image in which a part of the outline is lost. Character images 1301 and 1302 shown in FIG. 13 are both handwritten character images of the character “composition”. On the other hand, character images 1303 and 2304 are contour images extracted from the character images 1301 and 1302, respectively. In the character image 1302, a part of the outline has disappeared due to the collapse of the character. In such a case, the contour feature amount moment becomes unstable.

  A typical example of the present invention is as follows. That is, a character recognition device having an arithmetic device including a processor and a storage device, an input device connected to the arithmetic device, and an output device connected to the arithmetic device, wherein the arithmetic device includes the input device A first procedure for executing preprocessing for reducing disturbance factors that hinder character recognition from an input image input via a device or an input image stored in the storage device, and the preprocessing is executed. A second procedure for normalizing the image, a third procedure for converting the normalized image into a vector value on a vector space, and the vector value based on a recognition dictionary stored in the storage device. And a fifth procedure for outputting the result of the determination via the output device, wherein the second procedure is a contour of the image on which the preprocessing has been executed. Extract the first A procedure, a seventh procedure for synthesizing the preprocessed image and the extracted contour image, and a mapping from the synthesized image to a normalized image of a predetermined size. An eighth procedure for generating an image so that an image of the center of gravity of the generated image approaches the center of the range of the predetermined size, and an image of the range in which the pixels of the synthesized image spread approaches the range of the predetermined size; And a ninth procedure for normalizing the image on which the preprocessing has been performed according to the generated mapping.

  According to one embodiment of the present invention, normalization is performed based on a composite image of a character contour image and an original image, thereby reducing normalization instability when the character contour is lost, The recognition rate in handwritten characters can be improved.

It is a block diagram which shows an example of the hardware constitutions of the character recognition apparatus of embodiment of this invention. It is a flowchart which shows the process which the conventional character recognition apparatus performs. It is a flowchart which shows the outline | summary of the character recognition process performed by the arithmetic unit of embodiment of this invention. It is explanatory drawing of the identification process performed by the arithmetic unit of embodiment of this invention. It is explanatory drawing of the example of the image produced | generated by the moment normalization method and the bi-moment normalization method. It is explanatory drawing of the example of a different font of the same character. It is explanatory drawing of the gravity center and boundary of a character image defined by the moment normalization method. It is explanatory drawing of the 1st example of the extraction method of the character outline used by the calculating device of embodiment of this invention. It is explanatory drawing of the pixel referred in order to extract the outline of a character image in embodiment of this invention. It is explanatory drawing of the image normalized by the moment normalization method and the outline feature-value moment normalization method. It is explanatory drawing of the filter used in order to extract the outline of a character image in embodiment of this invention. It is explanatory drawing of the example of the outline image extracted by the arithmetic unit of embodiment of this invention. It is explanatory drawing which shows the example of the character from which a part of outline was lost.

  FIG. 1 is a block diagram showing an example of a hardware configuration of a character recognition device according to an embodiment of the present invention.

  The character recognition device 100 of the present invention includes an input device 101, an arithmetic device 102, a recognition dictionary 103, a display device 104, and a pattern database (DB) 105.

  The input device 101 is a device such as a keyboard and mouse for inputting commands and a scanner for inputting images.

  The arithmetic device 102 reads the input image and determines characters in the input image. The arithmetic device 102 includes a CPU (Central Processing Unit), a memory, a storage device, and the like.

  The recognition dictionary 103 is a dictionary database that stores a recognition dictionary.

  The display device 104 is a device that outputs the processing content of the arithmetic device 102, and is a device such as a display, for example. When it is not necessary to display the processing contents, the display device 104 may be omitted, and may be replaced by an output device other than the display device as necessary.

  The pattern DB 105 stores the pattern input by the input device 101.

  The recognition dictionary 103 and the pattern DB 105 may be stored in a storage device in the arithmetic device 102.

  The arithmetic device 102 according to the embodiment of the present invention includes character recognition means. Specifically, for example, the CPU in the arithmetic device 102 executes a program stored in a memory or a storage device, thereby realizing character recognition means.

  Next, the processing flow in the embodiment of the present invention will be described.

  FIG. 3 is a flowchart showing an outline of character recognition processing executed by the arithmetic device 102 according to the embodiment of this invention.

  The character image input unit 201, the preprocessing unit 202, the normalization unit 301, the feature extraction unit 209, the identification unit 210, the output unit 211, and the recognition dictionary learning unit 213 illustrated in FIG. (In other words, each program corresponds to a processing procedure executed by the arithmetic device 102). The same applies to the character outline extraction unit 302, the composite image generation unit 303, the moment value calculation unit 304, the normalized map generation unit 207, and the normalized image generation unit 208 included in the normalization unit 301.

  The character recognition device 100 reads an input image, determines a character in the input image, and outputs a determination result. As described above, FIG. 2 is a flowchart of character recognition processing by the conventional contour feature amount moment normalization method. Among the character recognition processes executed by the character recognition apparatus 100 of the present embodiment, the processes of the character outline extraction unit 302 and the composite image generation unit 303 in the normalization unit 301 are different from the conventional character recognition processing.

  In the character image input unit 201, an image to be recognized is input by a user or a program executed by the arithmetic device 102. For example, a scanner included in the input device 101 may read a document, and character data obtained by the scanner may be stored in a memory or a storage device. When character image data is stored in advance in a storage device or the like, it can be used as a recognition target.

  The preprocessing unit 202 performs noise removal, blurring processing, and the like on the input image, thereby reducing disturbance factors that hinder the determination of characters in the image, such as noise or blur. For example, an isolated point having a size equal to or smaller than a certain threshold is removed by noise removal processing. The input image that has undergone preprocessing may be temporarily stored in a storage device.

  The normalization unit 301 converts each pre-processed input image into a fixed-size image designated in advance. The converted image is called a normalized image. One of the main purposes of normalization is to unify subsequent processing by converting input images of various sizes into fixed-size images. Another main purpose of normalization is to convert various types of input images into fixed-size images so that variations in character shapes between the same characters are reduced. As a result, variations between character images of the same character type can be reduced, which helps to improve the identification accuracy. Details will be described later. The normalized image generated by the normalizing unit 301 may be temporarily stored in the storage device.

  The feature extraction unit 209 receives the normalized image generated by the normalization unit 301 as an input, and converts the input normalized image into a vector value on a vector space. The destination vector space is called a feature space, and the converted vector value is called a feature vector. Dimensional compression may reduce the dimension of the feature space. In that case, components having a small contribution to the identification are removed from the feature space as much as possible, and the feature vector is expressed as a feature vector on a lower-dimensional feature space.

  The identification unit 210 uses the recognition dictionary 214 to determine the character category to which the feature vector belongs. The recognition dictionary 214 holds information for dividing the feature space into regions occupied by each character category. As a result, the character category corresponding to the region to which the feature vector belongs is returned as the determination result.

  FIG. 4 is an explanatory diagram of identification processing executed by the arithmetic device 102 according to the embodiment of this invention.

  FIG. 4 shows, as an example, regions 402A, 402B, and 402C occupied by category A, category B, and category C in the feature space 401, respectively. Each category corresponds to one character. In this example, the unknown input (that is, the feature vector of the input normalized image) 403 is not included in any category area. In this case, the identification unit 210 may determine that the category A corresponding to the region 402A closest to the unknown input 403 is a category to which the unknown input belongs. Alternatively, the identification unit 210 may determine that the unknown input 403 does not belong to any category and determine rejection. The identification unit 210 outputs a determination result (for example, “category A” or “rejection”).

  Refer to FIG. 3 again. The output unit 211 outputs the determination result by the identification unit 210 to the display device 104 such as a display or a storage device.

  Next, before proceeding to the description of the processing of the normalization unit 301 of the present invention, the processing of the normalization unit 203 by the contour feature amount moment normalization method will be described.

  The size of the original image f (x, y) input to the character outline extraction unit 204 via the preprocessing unit 202 is assumed to be a width W0 and a height H0. Here, 0 ≦ x <W0, 0 ≦ y <H0, x and y representing each grid point are integer values, and the pixel values of the k1st grid point from the left and the k2th grid point from the bottom are f (k1-1, k2-1) is displayed. An example in which this original image is normalized to an image size having a width L and a height L will be described.

  When the contour feature amount moment normalization method is applied, first, the character contour extraction unit 204 extracts the character contour image fc (x, y) of the original image f (x, y). Two examples of contour extraction methods are given below.

  A first example of a character outline extraction method will be described. First, a horizontal component fx (x, y) and a vertical component fy (x, y) of the contour are extracted from the character image f (x, y).

  FIG. 8 is an explanatory diagram of a first example of a character outline extraction method used by the arithmetic device according to the embodiment of the present invention.

  FIG. 8 shows an input image 801, a contour image 802, a horizontal contour image 803, and a vertical contour image 804 as an example. Here, the input image 801 is an image of the letter “B”, and the contour image 802, the horizontal contour image 803, and the vertical contour image 804 are all examples of contour images extracted from the input image 801. is there. The input image 801 corresponds to f (x, y), the horizontal contour image 803 corresponds to fx (x, y), and the vertical contour image 804 corresponds to fy (x, y).

  First, the character outline extraction unit 204 sets fx (x, y) = 0 and fy (x, y) = 0. Next, the character outline extraction unit 204 sequentially selects the grid points of the input image f (x, y), and extracts the feature in the outline direction for each grid point. The diagonal direction is counted both vertically and horizontally. If the pixel at the grid point of interest is a black pixel, that is, p = f (x, y) = 1, the character outline extraction unit 204 stores information on pixels in the vicinity of the pixel p shown in FIG. From (1) to (3), features are extracted.

  FIG. 9 is an explanatory diagram of pixels referred to in order to extract the outline of the character image in the embodiment of the present invention.

  Specifically, FIG. 9 shows a positional relationship 901 between a pixel p at a certain grid point and pixels d1 to d7 at neighboring grid points adjacent thereto. For example, when the coordinates of the lattice points of the pixel p are (x, y), the coordinates of the lattice points of the pixels d1, d2, d3, d4, d5, d6, and d7 are (x + 1, y + 1), (x, y, respectively). y + 1), (x-1, y + 1), (x-1, y), (x-1, y-1), (x, y-1) and (x + 1, y-1).

  As a result, a horizontal component fx (x, y) and a vertical component fy (x, y) of the contour are generated. The contour image fc (x, y) is obtained by calculating fc (x, y) = fx (x, y) + fy (x, y) (= is substituted here).

  A second example of the character outline extraction method will be described. First, the character outline extraction unit 204 sets fx (x, y) = 0 and fy (x, y) = 0. Next, the character outline extraction unit 204 sequentially selects grid points of the input image f (x, y), and extracts features for each grid point. The diagonal direction is counted both vertically and horizontally. The character outline extraction unit 204 extracts features from the lattice point (x, y) according to Equation (4).

  FIG. 11 is an explanatory diagram of a filter used for extracting a contour of a character image in the embodiment of the present invention. The filter 1101 in FIG. 11 corresponds to the calculation formula of fy (x, y) in Expression (4), and the filter 1102 corresponds to the calculation expression of fx (x, y).

  The contour image fc (x, y) is obtained by calculating fc (x, y) = fx (x, y) + fy (x, y) (= is substituted here).

  FIG. 12 is an explanatory diagram illustrating an example of a contour image extracted by the arithmetic device 102 according to the embodiment of this invention. For example, the contour image 1202 is extracted from the original image 1201 of handwritten characters of the character types “Aya”, “鮎”, “or”, “粟”, and “袷”.

  Next, the moment value calculation unit 206 calculates the moment value of the contour image fc (x, y). Here, using the equations (5) and (7), the center of gravity (xc, yc) shown in the equation (6) and the values of δx and δy shown in the equation (8) are calculated. These δx and δy are parameters indicating the range in which the pixels of the original image spread, and are used to determine the boundary of the original image, which will be described later.

  Next, the normalized map generation unit 207 generates a map for mapping the original image to the normalized plane [0, L] × [0, L]. In the contour feature amount moment normalization method, an area having a width of δx in the horizontal direction and δy in the vertical direction around the center of gravity (xc, yc) calculated by the moment value calculation unit 206 is enlarged or reduced to L × By setting the size to L, a normalized image is generated. That is, a portion of [xc−δx / 2, xc + δx / 2] × [yc−δy / 2, yc + δy / 2] in the original image is mapped to the normalized plane [0, L] × [0, L]. The mapping for that is expressed by Equation (9).

  Next, the normalized image generation unit 208 generates a normalized image f ′ (x ′, y ′) by the relational expression of Expression (10). In the case of this example, as described above, the portion of [xc−δx / 2, xc + δx / 2] × [yc−δy / 2, yc + δy / 2] in the original image is enlarged to the size of L × L. By normalizing, a normalized image is obtained.

  As described above, the contour feature amount moment normalization method extracts the contour image fc (x, y) from the original image f (x, y), and uses the moment of the contour image fc (x, y) to Define the center of gravity and boundaries of the image.

  When a moment normalization method, i.e., a normalization method based on the moment value of the original image itself is applied as in the past, the image of the center of gravity of the pixel of the original image approaches the center of the range of the normalized image, In addition, a mapping from the original image to the normalized image is generated so that the image in the range where the pixels of the original image spread approaches the range of the normalized image.

  FIG. 7 is an explanatory diagram of the center of gravity and boundary of the character image defined by the moment normalization method.

  Specifically, FIG. 7 shows a pre-processed image (that is, the original image in the above description) 701, and a character image 702 including a center of gravity and a boundary defined for them. For example, the center of gravity 703A and the boundary 704A are determined from the original image 701A corresponding to the character type “0”. Here, the boundary 704A is a boundary between the area in which the character corresponding to the character type “0” is displayed and the other region in the original image 701A. In other words, the pixel of the character corresponding to the character type “0” Corresponds to a wide range. When the moment normalization method is applied, the secondary moment values δx and δy calculated by the equation (8) are used as parameters indicating the range in which the pixel of the character spreads, and the boundary 704A is laterally centered on the center of gravity 703A. It is defined by a region having a width of δx in the direction and δy in the vertical direction.

  Even if there are variations in the size and shape of the input character image by performing normalization using the mapping generated as described above, if they are images of the same character type, It can be expected to suppress variation in the feature amount of the normalized character image.

  However, when the moment normalization method as described above is applied, variations in the normalized image as shown in the normalized image 1002 in FIG. 10 occur according to the variation in the line thickness of the input character image. Cheap. This is because the moment value becomes unstable because the position of the center of gravity of the pixel of the original image fluctuates due to the influence of the thickness of the character line of the original image, and the mapping generated thereby fluctuates. It is to do.

  On the other hand, when the contour feature value moment normalization method is applied (that is, when normalization based on the moment value of the contour image is performed), the image of the center of gravity of the pixel of the contour of the original image is normalized. A map from the original image to the normalized image is generated so that the image in the range in which the pixels of the contour of the original image expand and approaches the center of the range of the original image approaches the range of the normalized image. In this case, since the pixels other than the outline are deleted from the original image, the position of the center of gravity of the outline pixel is hardly affected by the thickness of the character line of the original image. For this reason, the moment value and the generated mapping are stable regardless of the thickness of the character line, and as shown in the normalized image 1003 in FIG.

  However, as shown in the example of the character images 1302 and 1304 in FIG. 13, when the character outline structure is lost, a part of the outline cannot be extracted. Since the position of the center of gravity of the pixel of the contour fluctuates due to the loss of a part of the contour, the calculated moment value becomes unstable when the contour structure is lost, and the generated normalized image The variation increases between the same character types. Such variation appears as variation of vector points on the feature space after feature extraction, and causes a reduction in the identification rate.

  Next, normalization performed by the normalization unit 301 according to the embodiment of this invention will be described.

  The character outline extracting unit 302 may extract the character outline image fc (x, y) using the same method as the character outline extracting unit 204 (for example, the first or second example described above) A method may be used. Here, third and fourth examples will be described as examples of another method for extracting a character outline.

  First, a third example will be described. First, the character outline extraction unit 302 sets g0 (p) = g1 (p) =... = G7 (p) for all white pixels p = (x, y). Next, the character outline extraction unit 302 calculates g0 (p), g1 (p),..., G7 (p) with respect to all black pixels p = (x, y) by Expression (11).

  d0, d1,..., d7 are neighboring pixels of the pixel p as shown in FIG. The character outline extraction unit 302 generates an outline image fc (x, y) by fc (x, y) = Σgk (x, y). Here, Σgk (x, y) is calculated in the range of k = 0, 1,.

  Next, a fourth example will be described. First, the character outline extraction unit 302 calculates gx (p) and gy (p) for all the pixels p = (x, y) by using Expression (12). Next, the character outline extraction unit 302 generates an outline image fc (x, y) by using Equation (13). Here, d0, d1,..., D7 are neighboring pixels of the pixel p as shown in FIG.

  The above first to fourth examples are examples of methods for extracting the contour of a character image, and the character contour extraction unit 302 may extract the contour of a character image by a method other than the method exemplified above. . As described above, when the pixel values of the lattice points around the lattice points of the original image satisfy a predetermined condition, the pixel value of the contour image at the lattice points of the original image is increased (the first example described above) And equivalent to the third example), or the pixel values of the contour image at the grid points of the original image by summing the pixel values of the grid points around the grid points of the original image multiplied by a predetermined coefficient The outline of the character image can be extracted by the method of calculating the value (corresponding to the second example and the third example above).

  This is the end of the description of the character outline extraction unit 302. Next, processing after the composite image generation unit 303 will be described. The composite image generation unit 303 includes the character outline image fc (x, y) of each grid point generated by the character outline extraction unit 302 and the original image f (x, y) of each grid point output from the preprocessing unit 202. And a composite image fs (x, y) is generated by Expression (14).

  Here, γ1 and γ2 are positive numbers and satisfy γ1 + γ2 = 1. This composite image corresponds to an image in which the contour portion of the original image is emphasized, in other words, an image weighted so that the pixel value of the contour portion of the original image is larger than the pixel value of other portions.

  The moment value calculation unit 304 calculates a moment value using fs (x, y) instead of fc (x, y). That is, the moment value calculation unit 206 uses the formula (15) instead of the formula (5) to calculate the values of the center of gravity (xc, yc) shown in the formula (6) and δx and δy shown in the formula (8). .

  Subsequently, the normalized map generation unit 207 of the present invention generates a normalized map based on the moment value calculated by Equation (15) or the like, and uses the generated normalized map to normalize the present invention. The image generation unit 208 generates a normalized image (Formula (10)).

  In the embodiment of the present invention described above, a range (that is, a boundary) in which the pixels of the composite image spread is determined based on the secondary moment values δx and δy of the composite image. This range does not necessarily match the outline of the pixel of the composite image. However, the method of determining the range based on the moment value as described above is merely an example, and in the present invention, the range in which the pixels of the character image spread may be determined by a method other than the above. For example, instead of calculating the moment value in the moment value calculation unit 304, the arithmetic unit 102 may determine a rectangular range circumscribing the outline of the pixel of the composite image as a range in which the pixel of the composite image extends.

  Although the description so far relates to the character recognition device 100, the character recognition device 100 can also be used as a recognition dictionary generation device. In that case, the storage device of the arithmetic device 102 holds the character image DB 212 (FIG. 3), and the preprocessing unit 202 performs preprocessing on the character image stored in the character image DB 212. The processes of the normalization unit 301 and the feature extraction unit 209 are the same as those of the character recognition device 100 described above. The recognition dictionary learning unit 213 learns the recognition dictionary based on the feature amount extracted by the feature extraction unit 209 and stores the result in the recognition dictionary 214 (corresponding to the recognition dictionary 103 in FIG. 1). Note that the recognition dictionary learning unit 213 is a function realized by the arithmetic device 102 as in the normalization unit 301 and the like.

  As described above, according to the embodiment of the present invention, normalization is performed based on the moment value of the composite image of the original image and the contour image. That is, the moment value of the composite image is calculated, and based on this, a mapping from the original image to the normalized image is generated. By the synthesis, the pixel value of the outline portion of the character image becomes larger than the pixel value of the other portion. As a result, compared to the case where normalization based on the moment value of the original image itself is performed, the weight of the pixel in the contour portion is increased, so the influence of the thickness of the character line can be reduced, and Compared with the case where normalization based on the moment value of the contour image is performed, since the pixels other than the contour are also used, the influence of disappearance of the contour can be reduced. Thus, according to the present embodiment, stable normalization can be realized for both the thickness of the line and the disappearance of the outline, thereby improving the recognition rate of type characters and handwritten characters. Can do.

  In order to maximize the above effect, it is desirable to optimize the coefficients γ1 and γ2. The optimum values of the coefficients γ1 and γ2 may depend on various conditions such as the contour extraction method, but the pixel value of the contour portion of the character image in the synthesized image is larger than the pixel values of the other portions. It is necessary to select a correct value. For example, the composite image generation unit 303 of the present embodiment may use γ1 and γ2 that satisfy γ1 <γ2.

100 Character recognition device 101 Input device 102 Arithmetic devices 103, 214 Recognition dictionary 104 Display device 105 Pattern DB
201 Character image input unit 202 Pre-processing unit 203, 301 Normalization unit 204, 302 Character contour extraction unit 206, 304 Moment value calculation unit 207 Normalization map generation unit 208 Normalized image generation unit 209 Feature extraction unit 210 Identification unit 211 Output Part 212 Character Image DB
213 Recognition dictionary learning unit 303 Composite image generation unit

Claims (15)

A character recognition device having an arithmetic device including a processor and a storage device, an input device connected to the arithmetic device, and an output device connected to the arithmetic device,
The arithmetic unit is:
A first procedure for performing preprocessing for reducing disturbance factors that hinder character recognition from an input image input via the input device or an input image stored in the storage device;
A second procedure for normalizing the preprocessed image;
A third procedure for converting the normalized image into a vector value on a vector space;
A fourth procedure for determining that the vector value is any character based on a recognition dictionary stored in the storage device;
Executing a fifth procedure for outputting the result of the determination via the output device;
The second procedure includes
A sixth procedure for extracting the contour of the image on which the preprocessing has been performed;
A seventh procedure for synthesizing the preprocessed image and the extracted contour image;
Mapping the synthesized image from the synthesized image to a normalized image of a predetermined size, wherein the image of the center of gravity of the synthesized image approaches the center of the range of the predetermined size, and the pixel of the synthesized image An eighth procedure for generating an image in a range in which the image expands so as to approach the range of the predetermined size;
And a ninth procedure for normalizing the image on which the preprocessing has been performed according to the generated mapping.   In the seventh procedure, a value obtained by multiplying the pixel value of the image on which the preprocessing is performed at each grid point by the first coefficient, and a pixel value of the contour image at each grid point are multiplied by the second coefficient. The character recognition apparatus according to claim 1, further comprising: calculating a pixel value of the synthesized image at each lattice point by adding the calculated values. The second procedure further includes a step of calculating a moment value of the synthesized image as a parameter indicating a range in which pixels of the synthesized image are spread,
The character recognition device according to claim 1, wherein the eighth procedure includes a procedure of generating a map for enlarging or reducing the synthesized image in accordance with the moment value.   The sixth procedure includes a step of increasing the pixel value of the contour image at each grid point when the pixel values of the grid points around each grid point of the image on which the preprocessing has been executed satisfy a predetermined condition. The character recognition device according to claim 1, further comprising:   In the sixth procedure, the pixel value of the contour image at each grid point is calculated by aggregating values obtained by multiplying pixel values of grid points around each grid point of the preprocessed image by a predetermined coefficient. The character recognition device according to claim 1, further comprising: A recognition dictionary generation device having a processor and an arithmetic device connected to the processor and including a storage device storing a character image,
The arithmetic unit is:
A first procedure for performing preprocessing for reducing disturbance factors that hinder character recognition from the character image stored in the storage device;
A second procedure for normalizing the preprocessed image;
A third procedure for converting the normalized image into a vector value on a vector space;
A fourth procedure for learning a recognition dictionary used for character recognition based on the vector value;
And a fifth procedure for storing the learning result in the storage device,
The second procedure includes
A sixth procedure for extracting the contour of the image on which the preprocessing has been performed;
A seventh procedure for synthesizing the preprocessed image and the extracted contour image;
Mapping the synthesized image from the synthesized image to a normalized image of a predetermined size, wherein the image of the center of gravity of the synthesized image approaches the center of the range of the predetermined size, and the pixel of the synthesized image An eighth procedure for generating an image in a range in which the image expands so as to approach the range of the predetermined size;
And a ninth procedure for normalizing the image on which the preprocessing has been performed according to the generated mapping.   In the seventh procedure, a value obtained by multiplying the pixel value of the image on which the preprocessing is performed at each grid point by the first coefficient, and a pixel value of the contour image at each grid point are multiplied by the second coefficient. The recognition dictionary generation device according to claim 6, further comprising: calculating a pixel value of the synthesized image at each lattice point by adding the obtained values. The second procedure further includes a step of calculating a moment value of the synthesized image as a parameter indicating a range in which pixels of the synthesized image are spread,
The recognition dictionary generating apparatus according to claim 6, wherein the eighth procedure includes a procedure of generating a map for enlarging or reducing the synthesized image in accordance with the moment value.   The sixth procedure includes a step of increasing the pixel value of the contour image at each grid point when the pixel values of the grid points around each grid point of the image on which the preprocessing has been executed satisfy a predetermined condition. The recognition dictionary generation device according to claim 6, further comprising:   In the sixth procedure, the pixel value of the contour image at each grid point is calculated by aggregating values obtained by multiplying pixel values of grid points around each grid point of the preprocessed image by a predetermined coefficient. The recognition dictionary generation device according to claim 6, further comprising: An image normalization method executed by an arithmetic device including a processor and a storage device connected to the processor,
A sixth procedure in which the arithmetic device extracts the contour of the original image stored in the storage device;
A seventh procedure in which the arithmetic device synthesizes the original image and the extracted contour image;
The computing device maps the synthesized image to a normalized image of a predetermined size, and the image of the center of gravity of the synthesized image approaches the center of the range of the predetermined size, and the synthesis An eighth procedure for generating an image in a range in which pixels of the generated image expand so as to approach the range of the predetermined size;
A normalization method comprising: a ninth procedure in which the arithmetic device normalizes the original image according to the generated mapping and stores the result in the storage device.   In the seventh procedure, a value obtained by multiplying the pixel value of the image on which the preprocessing is performed at each grid point by the first coefficient, and a pixel value of the contour image at each grid point are multiplied by the second coefficient. The normalization method according to claim 11, further comprising: calculating a pixel value of the synthesized image at each grid point by adding the calculated values. The normalization method further includes a procedure of calculating a moment value of the synthesized image as a parameter indicating a range in which pixels of the synthesized image are spread,
The normalization method according to claim 11, wherein the eighth procedure includes a procedure of generating a map for enlarging or reducing the synthesized image in accordance with the moment value.   The sixth procedure includes a step of increasing the pixel value of the contour image at each grid point when the pixel values of the grid points around each grid point of the image on which the preprocessing has been executed satisfy a predetermined condition. The normalization method according to claim 11, further comprising:   In the sixth procedure, the pixel value of the contour image at each grid point is calculated by aggregating values obtained by multiplying pixel values of grid points around each grid point of the preprocessed image by a predetermined coefficient. The normalization method according to claim 11, further comprising a step of calculating

Images