JP2005063056A - Image recognition device, image recognizing program, and recording medium with the program stored therein - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To generate reproduced images, in which characters and diagrams are accurately reproduced or improved from an image of a document and a drawing which have been captured by an image input device. <P>SOLUTION: An image recognition device is provided with an original image acquiring means 201; a high-resolution multi-level image generating means 202 for generating a high-resolution multi-level image from an original image; a brightness curved surface generating means 204 for generating a curved surface whose xy coordinates are pixel coordinates and whose z coordinate indicates brightness; a topographic feature extracting means 205 for extracting a topographic feature of the brightness curved surface; a topographic feature incorporating means 206 for generating a binary image, in which the characters and diagrams are reproduced by incorporating the topographic feature into the contours of the characters and diagrams included in the multi-level image generated by the high-resolution multi-level image generating means; contour correction means 2081, 2082 for correcting the contours of the characters and diagrams included in the binary image generated by the topographic feature incorporating means; and a character/diagram recognition means 209 for recognizing the characters and diagrams in the image as character codes and vector diagrams. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention creates an image that accurately reproduces a character / line diagram or an image that is reproduced by improving the character / line diagram from a document or drawing image captured by an image input device such as a scanner or a digital camera, The present invention relates to an image recognition apparatus capable of recognizing the character / line diagram as a character code / vector diagram, an image recognition program, and a recording medium on which the program is recorded.

  In recent years, color information input devices such as a color scanner and a color digital camera (including a mobile phone with a color digital camera) are widely used. For example, when a document image is read at a resolution of 400 dpi and 256 gradations, the color scanner has a disadvantage that the amount of data to be processed increases and it takes time to read. In addition, the color digital camera has a disadvantage that the number of images that can be stored in the memory is reduced, for example, when shooting at a resolution of 400 dpi.

  In general, even in the case of a color image, for example, about 200 dpi, the image quality does not look inferior when displayed on a display. For this reason, in this type of color information input device, color documents, drawings, maps, etc. are acquired (that is, read or photographed) with a resolution of about 200 dpi, and an image file in a predetermined format (TIF, JPEG, GIF, etc.). To other devices.

  By the way, when a color bitmap image includes characters and diagrams in addition to photographs, pictures, etc., there are cases where it is desired to recognize these sentences and diagrams accurately or with improvement. For example, a user may want to process a character / line diagram drawn in a color bitmap image with a graphic tool (software that can process an image). Alternatively, there is a case where a user wants to edit a character string drawn in a color bitmap image by a character recognition device and edit it as a character string that can be edited by a word processor.

  However, when a color print document, drawing, map, etc. are created with a resolution of about 200 dpi, if the character / line diagram included in the image is small (for example, about 3 mm square), it is included in the image. When a character / diagram is recognized by a recognition device such as OCR, the character / diagram may not be reproduced correctly.

  Documents, drawings, maps, etc. composed of monochrome images appear inferior in image quality when displayed on a display, for example, at a resolution of about 200 dpi. For this reason, a document or the like composed of a monochrome image is usually created with a resolution of 400 dpi or higher. That is, with a monochrome scanner (or in a monochrome reading mode of a color scanner), a document, drawing, map, etc. can be read at a resolution of about 400 dpi and passed to a computer as a binary image. Also, with a digital camera, when shooting in monochrome, the amount of captured image data is overwhelmingly smaller than the amount of data captured in color, so the captured image is stored in memory with a resolution of about 400 dpi. The

  If the monochrome bitmap image has a resolution of about 400 dpi, even if characters in the image are small (for example, about 3 mm square), for example, recognition by a character recognition device can be performed well.

  Conventionally, a diagram or the like included in a multi-tone color image having a resolution of about 200 dpi or less is accurately reproduced or improved in the same manner as monochrome 400 dpi, and this is recognized as, for example, a character. It is desired.

  To accurately reproduce or improve the characters drawn in a conventional color bitmap image, (1) filtering, (2) contrast enhancement, (3) model-based image restoration, (4 ) Higher resolution is used.

  For the filtering of (1), for example, a noise removal method using Morphologic (see Non-Patent Document 1 and Non-Patent Document 2), a method using a secondary filter that removes noise without blurring details ( Non-Patent Document 3).

  For the contrast enhancement in (2), for example, nonlinear tone conversion based on local statistics within a processing window (for example, a processing region of 3 × 3 pixels and 5 × 5 pixels) (see Non-Patent Document 4) )

  In model-based image restoration (3), there is a method of modeling the cause of misrecognition such as OCR by cluster analysis and restoring the image (see Non-Patent Document 5).

  (4) As an increase in resolution, a method of generating an outline of an arbitrary resolution by clustering and averaging pixels constituting a character / line diagram in a bitmap image (see Non-Patent Document 6), distribution Based on an evaluation function with three parameters of bimodality, smoothness, and luminance as a parameter, it is formulated as an inverse problem, and an optimal high-resolution image is restored (see Non-Patent Document 7). There are methods based on binarization (see Patent Literature 1, Patent Literature 2, Patent Literature 3, Non-Patent Literature 8, and Non-Patent Literature 9).

When considering improvement of recognition accuracy in a character recognition device or the like, an approach of restoring a high-resolution binary bitmap image from a low-resolution bitmap image is natural.
Theoretically, the sampling interval, the number of quantization levels, and the PSF (point spread function) are blurred between a low-resolution / multi-gradation (or multi-value) image and a high-resolution / binary image. It is known that the same amount of information is obtained if a certain relationship is established regarding the effect (Non-patent Document 10).

Japanese Patent No. 3345350 JP-A-8-340446 JP 2001-118032 A L. Koskinen, H. Huttunen, and J.T.Astola, Text enhancement method based on soft morphological filters Proceedings of SPIE, vol. 2181, pp. 243-253, 1994. J. Liang, R.M.Halalick, and I.T.Phillips, Document image restoration using binary morphological filters, Proceedings of SPIE, vol. 2660, pp. 274-285, 1996. G. Ramponi and P. Fontanot, Enhancing document images with a quadratic filter, Signal Processing, vol. 33, pp. 23-34. 1993. Y.C. Shin, R. Sridhar, V. Demjanenko, P.W.Palumbo, and J.J.Hull, Contrast enhancement of mail piece images, Proceedings of SPIE, vol. 1661, pp. 27-37, 1992. M.Y.Jaisimha, E.A.Riskin, R. Ladner, and S. Werner, Model-based restoration of document images for OCR, Proceedings of SPIE, vol. 2660, pp. 297-308, 1996.1997 US5930393 (T.K.Ho and J.D.Hobby; Lucent Technologies), July 27, 1999. P.D.Thouin and C.-I.Chang, A method for restoration of low-resolution document images, International Journal on Document Analysis and Recognition, vol. 2, pp. 200-210, 2000. US5524070 (Y.-C. Shin, R. Sridhar, S.N.Srihari and V. Demjamenko; State University of New York, Buffalo), April 6, 1996. US6347156 (H. Kamada and K. Fujimoto; Fujitsu), February 12, 2002. D. Lee, T. Pavlidis, and G.W. Wasilkowski, A note on the trade-off between sampling and quantization in signal processing, Journal of Complexity, vol. 3, pp. 359-371, 1987. N.B.Karayiannis and A.N.Venetsanopoulos, Image interpolation based on variational principles, Signal Processing, vol. 25, pp. 259-288, 1991. A.D.Kulkarni and K. Sivaraman, Interpolation of digital imagery using hyperspace approximation, Signal Processing, vol. 7, pp. 65-73, 1984.

However, the conventional method described above has the following problems.
First, in the bitmap clustering and averaging method (1), it is assumed that a sufficient number of samples exist for the same character on the same document image. This assumption is valid when the number of character types is small as in European languages (about 100 at most including alphabets, numbers, symbols, etc.), but when there are many character types as in Eastern languages In many cases (for example, several thousand to several tens of thousands in the case of kanji) is not established.

  Next, there is a method of formulating and solving as the inverse problem of (4). However, when the calculation amount is large and the stroke density is high like kanji, FIG. 12 (A), (B) described later. , (C), the brightness of the stroke and the background may be reversed, and it may be difficult to restore due to the bimodality of the distribution and the brightness. Even a simple method of interpolation and binarization can restore the font shape characteristics to some extent. However, as shown in FIG. 10B, which will be described later, it is easy for outlines to be rattled and strokes to be cut off, and sufficient image quality cannot be obtained.

  If the image has a low resolution (for example, about 200 dpi), the stroke width is about the same as the sampling interval for small characters (for example, 3 mm square characters). Therefore, since the variation in the brightness of the stroke exceeds the range that can be explained by statistical fluctuations, the stroke may not be extracted by normal binarization.

  If the parameters are adjusted to extract blurry (ie, blurry) strokes, this time collapse will occur. In addition to bilinear interpolation, cubic spline interpolation, etc., there are various types of interpolation such as those based on the variational principle (Non-Patent Document 11) and those based on orthogonal polynomial bases (Non-Patent Document 12). Although there are techniques, these interpolation techniques are effective for natural images (photos), but there are no techniques that are particularly effective for text and diagrams.

  Further, although the method of clustering (2) and (3) can finely correct the character outline, it cannot perform processing such as character stroke extraction.

  As a result, even if the character / diagram is modified, high recognition accuracy of the character / diagram cannot be obtained.

  An object of the present invention is not only a character with a small number of characters such as alphabets, but also a simple algorithm that has a large number of characters such as kanji and a complicated shape structure, and further shows a line diagram with a fine structure. An image recognition apparatus, an image recognition program, and an image recognition program capable of accurately recognizing characters and diagrams by converting them into character codes or vector figures and recognizing them with high accuracy It is to provide a recording medium on which a program is recorded.

[1] The present invention provides: “Original image acquisition means for acquiring an original image of a multilevel bitmap;
From the original image acquired by the original image acquisition means, a high-resolution multi-value image generation means for generating a multi-value image having a higher resolution than the original image;
A luminance curved surface generating means for generating a curved surface in which the xy coordinates are pixel coordinates and the z coordinate is luminance from the original image acquired by the original image acquiring means;
Topographic feature extraction means for extracting topographic features of the luminance curved surface generated by the luminance curved surface generation means;
The character / line diagram is reproduced by incorporating the topographic feature extracted by the topographic feature extracting unit into the outline of the character / line diagram included in the multi-level image generated by the high-resolution multi-level image generating unit. Or a topographic feature embedding means for generating an improved binary image;
Contour correcting means for correcting the contour of a character / line diagram included in the binary image generated by the topographic feature incorporating means;
A character / line diagram recognition unit for recognizing the character / line diagram as a character code / vector diagram of the image in which the contour of the character / line diagram is corrected by the contour correction unit;
An image recognition apparatus comprising: "
Is the gist.

[2] According to the present invention, “the topographical feature is a“ valley or depression ”whose brightness is lower than the surroundings when the brightness curved surface is made to correspond to the actual topography, and“ ridge or mountaintop ”whose brightness is higher than the surroundings. The image recognition apparatus according to claim 1, wherein the image recognition apparatus is a “mountain or ridge” located between a “valley or depression” and a “ridge or mountain peak”. "
Is the gist.
Specifically, the topographic feature can be expressed by setting “valley or depression” to black, “ridge or mountaintop” to white, and “mountain or ridge” to gray. In this case, “valleys and depressions”, “ridges and peaks”, “hillsides or ridges” are not determined by altitude (luminance value: absolute value), but altitude difference (luminance value difference: relative value) It is expressed by

[3] The present invention provides: “Furthermore, a color restoration means is provided,
The color restoration unit restores the character / line diagram recognized by the character / line diagram recognition unit or the background color by approximating the color of the original image acquired by the original image acquisition unit. The image recognition device according to [1] or [2]. "
Is the gist.

[4] The present invention is characterized in that "the original image acquisition means includes color / grayscale conversion means for converting the original image into a grayscale multilevel image when the original image is a color image. The image recognition device according to any one of [1] to [3].
Is the gist.

[5] According to the present invention, “the topographic feature incorporation means includes a region of the luminance curved surface in which a minimal portion is continuous in a linear or belt shape and a region in which the minimal portion is localized. The image recognition apparatus according to any one of [1] to [4], wherein the binary image is generated by reproducing or improving a character / line diagram in the original image.
Is the gist.

[6] According to the present invention, (a) a character / line diagram is configured based on statistical information referring to surrounding pixels of each pixel from the high-resolution multi-value image generated by the high-resolution multi-value image generation unit. A basic image (an image in which topographic features are incorporated) composed of pixels, (b) pixels that do not constitute a character / line diagram, and (c) pixels that are not determined whether to constitute a character / line diagram. ) To generate basic image generation means,
The topographic feature incorporation means generates a binary image that reproduces or improves the character / line diagram based on the basic image generated by the basic image generation means. The image recognition apparatus according to any one of the above. "
Is the gist.

[7] According to the present invention, “the topographic feature incorporation means has (c) a pixel for which it is not determined whether or not to constitute a character / line diagram, a minimum portion of the luminance curved surface is linear or band-shaped. If the pixel is included in a region that is continuous with the region or the region where the local minimum portion is localized, the binary image in which the pixel / line diagram is reproduced or improved is assumed to constitute the character / line diagram. The image recognizing device according to [6], which is characterized.
Is the gist.

[8] The present invention includes: “a luminance gradient amount detection unit that detects a luminance gradient amount in each pixel of the high resolution multilevel image generated by the high resolution multilevel image generation unit;
The contour correcting unit reproduces or improves the character / line diagram created by the topographic feature incorporation unit based on the luminance gradient amount of each pixel detected by the luminance gradient amount detecting unit. First contour correcting means for correcting
The image recognition apparatus according to any one of [1] to [7], comprising: "
Is the gist.

[9] According to the present invention, “the contour correcting unit further refers to a curvature or direction change of a pixel row forming a contour of a character / line diagram in the image corrected by the first contour correcting unit. [8] The image recognition apparatus according to [8], further including second contour correcting means for correcting a contour of a character / line diagram in an image.
Is the gist.

[10] According to the present invention, “the second contour correcting unit clusters the pixel columns constituting the contour of the character / line diagram in the image corrected by the first contour correcting unit based on the tangent direction. Then, the image recognition apparatus according to [9], wherein the outline of the character / line diagram in the image is corrected by smoothing the outline and / or sharpening the corners of each cluster. "
Is the gist.

[11] The present invention provides: “Original image acquisition means for acquiring an original image of a multilevel bitmap;
From the original image acquired by the original image acquisition means, a high-resolution multi-value image generation means for generating a multi-value image having a higher resolution than the original image;
A luminance curved surface generating means for generating a curved surface in which the xy coordinates are pixel coordinates and the z coordinate is luminance from the multi-value image generated by the high-resolution multi-value image generating means;
Topographic feature extraction means for extracting topographic features of the luminance curved surface generated by the luminance curved surface generation means;
Topographic feature incorporation means for creating a binary image that reproduces or improves a character / line diagram in the original image by incorporating the topographic feature extracted by the topographic feature extraction means;
Contour correcting means for correcting the contour of a character / line diagram included in the binary image generated by the topographic feature incorporating means;
Contour correcting means for correcting the contour of a character / line diagram included in the binary image generated by the topographic feature incorporating means;
A character / line diagram recognition unit for recognizing the character / line diagram of the image whose contour has been corrected by the contour correction unit as a character code / vector diagram;
An image recognition apparatus comprising: "
Is the gist.

[12] According to the present invention, “the topographic feature is a“ valley or depression ”having a lower luminance than the surroundings when the luminance curved surface is made to correspond to the actual terrain, and a“ ridge or mountaintop ”having a higher luminance than the surroundings. [11] The image recognition device according to [11], wherein the image recognition device is a “hillside or ridge” located between a “valley or depression” and a “ridge or mountaintop”. "
Is the gist.

[13] According to the present invention, "the character / diagram or the background color of the image output by the character / diagram recognition unit is further approximated to the color in the original image acquired by the original image acquisition unit". The image recognition device according to [11] or [12], wherein
Is the gist.

[14] The present invention is characterized in that the original image acquisition means includes a color / grayscale conversion means for converting the original image into a grayscale multilevel image when the original image is a color image. The image recognition device according to any one of [11] to [13].
Is the gist.

[15] According to the present invention, “the topographic feature incorporation means includes a part of a character / diagram in which the minimum part of the luminance curved surface is continuous in a linear or belt-like form and the region in which the minimum part is localized. The image recognition according to any one of [11] to [14], wherein the binary image is generated by reproducing or improving a character / line diagram obtained by reproducing or improving a character / line diagram in the original image apparatus."
Is the gist.

[16] According to the present invention, “(a) a character / line diagram is constructed based on statistical information referring to surrounding pixels of each pixel from the high-resolution multi-value image generated by the high-resolution multi-value image generation means”. Basic image generation means for generating a basic image comprising pixels, (b) pixels that do not constitute a character / line diagram, and (c) pixels that are not determined whether to constitute a character / line diagram,
[11] to [15], wherein the topographic feature incorporation means generates a binary image in which the character / line diagram is reproduced or improved based on the basic image generated by the basic image generation means. The image recognition apparatus according to any one of the above. "
Is the gist.

[17] The present invention is as follows: "The topographic feature incorporation means has (c) a pixel for which it is not determined whether or not to constitute a character / line diagram, and a minimum portion of the luminance curved surface is linear or band-shaped. If the pixel is included in a region that is continuous with the region or the region where the local minimum portion is localized, the binary image in which the pixel / line diagram is reproduced or improved is assumed to constitute the character / line diagram. The image recognizing device according to [16], which is characterized.
Is the gist.

[18] The present invention includes: “a luminance gradient amount detection unit that detects a luminance gradient amount in each pixel of the high resolution multilevel image generated by the high resolution multilevel image generation unit;
The contour correcting unit reproduces or improves the character / line diagram created by the topographic feature incorporation unit based on the luminance gradient amount of each pixel detected by the luminance gradient amount detecting unit. First contour correcting means for correcting
The image recognition apparatus according to any one of [11] to [17], wherein "
Is the gist.

[19] The present invention provides: “The contour correcting means further refers to a curvature or direction change of a pixel row forming a contour of a character / line diagram in the image corrected by the first contour correcting means. The image recognition apparatus according to [18], further comprising second contour correcting means for correcting a contour of a character / line diagram in an image.
Is the gist.

[20] The second contour correcting unit clusters the pixel columns constituting the contour of the character / line diagram in the image corrected by the first contour correcting unit on the basis of the tangent direction. The image recognition apparatus according to [19], wherein the outline of the character / line diagram in the image is corrected by smoothing the outline and / or sharpening the corners. "
Is the gist.

[21] The present invention provides “an image recognition program that causes a computer to function as each means according to any one of [1] to [20].”
Is the gist.

[22] The present invention is “a computer-readable recording medium recording a program for causing a computer to be executed as each of the means according to any one of [1] to [20].”
Is the gist.

According to the present invention, it is possible to reproduce a character / line diagram having a complicated shape structure such as a Chinese character with high recognition accuracy. In addition, since processing is performed with simple calculations, characters and diagrams can be reproduced at high speed.
Image quality is improved in terms of text readability (especially smoothness, linearity, and stroke recovery of contours).

FIG. 1 is a diagram showing a configuration example of an image recognition apparatus according to the present invention.
In FIG. 1, a personal computer (PC) 110 operates as an image recognition device, and includes a CPU 111, a memory 112 (consisting of a ROM 1121 and a RAM 1122), a hard disk device 113, a removable disk device 114, and a display interface 115. A printer interface 116, a keyboard 117, and a network interface 118 connected to the communication line 100 are connected to a bus 119.

  The memory 112 includes an original image acquisition unit 201, an original high-resolution multilevel image generation unit 202, a basic image generation unit 203, a luminance curved surface generation unit 204, a topographic feature extraction unit 205, a topographic feature incorporation unit 206, a luminance, which will be described later. A program that functions as the gradient amount detecting means 207, the contour correcting means 208 (the first contour correcting means 2081, the second contour correcting means 2082), and the character / line diagram recognition means 209 is stored. These programs constitute the image recognition program of the present invention.

  Here, it is assumed that a color bitmap image that can be a target of image recognition is stored in the removable disk device 114. Further, it is assumed that the image recognition program of the present invention is stored in the hard disk device 113, and when the user starts the image recognition program, the image recognition program (indicated by “GRP”) is read into the RAM 1122 and image creation processing is possible. become.

In order to accurately reproduce the original image or to reproduce the original image with improvement, it is effective to use various information in various ways (top-down / bottom-up).
In the present invention, high-accuracy image recognition can be performed by adopting a method of correcting a basic image generated by interpolation and local statistics by integrating various features obtained by image analysis.
The process in this invention can be comprised combining the process of (2)-(5) with the process of (1).

(1) Interpolation of missing strokes by topographic features A luminance curved surface z = f (x, y) is formed from a processing target image (original image of a multi-value bitmap). The topographic features (ridge, canyon, summit, depression, hillside, ridge) of this luminance curved surface z = f (x, y) are examined. When the original image is a color bitmap image, the original image is converted into a grayscale multi-value bitmap image, which is used as a processing target image.

Comparing the height of each pixel with the surroundings on the luminance curved surface f (x, y), a stroke portion (that is, “valley or depression”) that is a series of pixels that are lower (darker) than the surroundings, than the surroundings. It can be classified into three levels: a portion corresponding to a gap between strokes that is a series of high (bright) pixels (ie, “ridge or peak”), and other (background: ie, “hillside or ridge”). FIG. 9 is a topographic feature map GioC in which each pixel is classified in this way by the luminance curved surface z = f (x, y). Here, four Japanese characters (C1: “skill” in the multi-value bitmap image are shown. , C2: “group”, C3: “still”, C4: “ka”).
When four Japanese characters (C1, C2, C3, C4) in a multilevel bitmap image are represented by three levels (high, medium, low), the multilevel bitmap image has a low resolution (for example, about 200 dpi). But the character features are preserved without loss.
Therefore, it is extremely effective to refer to the level of the luminance curved surface, particularly when a stroke is extracted from an image having a complicated structure such as the above-described Chinese character.

Therefore, in each of the embodiments described later, a stroke (see FIG. 10B) that is lost when the original image is binarized as it is by incorporating the topographic features calculated from the low-resolution original image. Complement.
For example, the “gorge” portion (black portion) of the topographic feature map GioC as shown in FIG. 9 is an area in which the minimal portion is continuous in a line shape or a belt shape, and corresponds to a “line segment with a long stroke”. Further, the “recess” portion is a region where the minimal portion is localized, and corresponds to “a line segment having a short point or stroke”.
If there are other “canyons” or “recesses” in the vicinity of the end of the “canyon”, the “canyons” should be connected to other “canyons” or “recesses”. If there is another “kubochi” or “gorge” in the vicinity of the end of “,” the “depression” should be continuous with the other “pit” or “canyon”. Therefore, the stroke can be created accurately or with improved by referring to the surrounding pixels of the “canyon” and “canyon”.
Wang et al. (L. Wang and T. Pavlidis, Direct gray-scale extraction of features for character recognition, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 15, no. 10, pp. 1053-1067, 1993.) As discussed in, it is particularly useful in low resolution images to consider topographic features.

(2) Interpolated image generation and basic image generation processing based on local statistics Considering the generality that the calculation is simple and the objects to be handled are not limited (that is, any character or diagram can be handled). Then, the following processing is performed.
(I) A high-resolution multilevel bitmap image can be generated from the original image by interpolation, and the missing stroke complementing process according to the topographic feature of (1) can be applied to this image.
(Ii) A basic image based on local statistics is generated from the original image, and the missing stroke complementing process based on the topographic features of (1) can be performed on the image. With the basic image, as shown in FIG. 11A, the luminance value of each pixel can take a value of three levels, and “ON” (pixels constituting a character / line diagram, here z = f ( x, y) = 1) is a black region, “OFF” (a pixel not constituting a character / line diagram, here z = f (x, y) = 0) is a white region, “TBD” ”(Pixels for which it is not determined whether or not to constitute a character / line diagram, here z = f (x, y) = − 1) is a gray region.
(Iii) A high-resolution multilevel bitmap image is generated from the original image by interpolation, and a basic image based on local statistics is generated from this image. Then, the missing stroke complementing process based on the topographic feature (1) can be performed on the basic image.

(3) First correction of outline Even if the missing strokes can be complemented by (1) and (2), the outline is not smooth, and the quality of the character image may be poor.
As a reference for a smooth contour line, it is necessary that the curvature is small (the radius of curvature is large). Further, it is assumed that the contour line passes through a pixel having a maximum luminance gradient amount (that is, a boundary between the white area and the black area).
Therefore, the Active Contour Model or Snake method (D. Williams and M. Shah, A fast algorithm for active contours and curvature estimation, CVGIP: Image Understanding, vol. 55 No. 1, pp. 14-26, 1992.) to correct the contour line (see the correction result in FIG. 20).

(4) Second correction of the contour In order to correct the shakiness of the horizontal and vertical lines and the dullness of the intersection of the vertical and horizontal lines, which are points of concern for human vision, Two corrections can also be made.
(5) Image recognition Character / diagram recognition is performed on images whose outlines have been restored or improved.

Hereinafter, embodiments of the image recognition apparatus of the present invention will be described.
<< First Embodiment >>
FIG. 2 shows the configuration of the image recognition apparatus 2A of the first embodiment.
In the first embodiment, the image recognition apparatus 2A includes an original image acquisition unit 201, an original high-resolution multilevel image generation unit 202, a basic image generation unit 203, a luminance curved surface generation unit 204, a topographic feature extraction unit 205, and a topographic feature. An incorporating unit 206, a luminance gradient amount detecting unit 207, a first contour correcting unit 2081, and a character / line diagram recognizing unit 209 are provided.
The original image acquisition unit 201 is an original image acquisition program, the original high resolution multilevel image generation unit 202 is a high resolution multilevel image generation program, the basic image generation unit 203 is a basic image generation program, and the luminance curved surface generation unit 204 is a luminance. According to the curved surface generation program, the topographic feature extraction means 205 is a topographic feature extraction program, the topographic feature incorporation means 206 is a topographic feature incorporation program, and the brightness gradient amount detection means 207 is a brightness gradient amount detection program. The contour correcting unit 2081 can be realized by the first contour correcting program, and the character / line diagram recognition unit 209 (pattern recognition unit 2091, pattern storage unit 2092) can be realized by the character / line diagram recognition program.
In FIG. 2, an image output unit 30 that outputs an image recognized by the character / line diagram recognition unit 209 is connected to the character / line diagram recognition unit 209.

The original image acquisition unit 201 can acquire an original image of a multi-value bitmap, and when this original image is in color, a color / gray scale conversion unit 2011 that converts the original image into a gray scale multi-value image. It has.
In the first embodiment, the original high-resolution multivalued image generation unit 202 can generate a multivalued image having a resolution higher than the resolution of the original image from the original image acquired by the original image extraction unit 11.
The basic image generation means 203 is based on statistical information referring to surrounding pixels of each pixel from the multi-value image generated by the original high-resolution multi-value image generation means 202, and (a) pixels constituting a character / line diagram; A basic image can be generated from (b) pixels that do not constitute a character / line diagram and (c) pixels that are not determined whether to constitute a character / line diagram.

The luminance curved surface generation unit 204 generates a curved surface z = f (x, y) from the original image acquired by the original image acquisition unit 201, where the xy coordinates are pixel coordinates and the z coordinate is luminance. For example, the luminance can take one of three values: high level, medium level, and low level. F (x, y) = 1 for high luminance and f (x, y for medium luminance. ) = 0, and in the case of low luminance, f (x, y) =-1.
The topographic feature extraction unit 205 can extract the topographic features (ridge, canyon, summit, depression, hillside) of the luminance curved surface f (x, y) generated by the luminance curved surface generation unit 204.

The topographical feature incorporation means 206 is a character / line diagram showing a region (canyon) in which the minimal portion continues linearly or in a strip shape and a region (concave) in which the minimal portion is localized in the luminance curved surface f (x, y). As a part, it is possible to generate a binary image in which a character / line diagram in the original image is reproduced or improved. For example, the topographic feature incorporation means 206 forms a character / line diagram from a high-resolution original image by an appropriate method (for example, an appropriate filter) and does not constitute the character / line diagram, but the luminance curved surface f ( The pixels corresponding to the canyons and depressions of x, y) can be modified to form a character / line diagram.
Then, the topographic feature incorporation means 206 (c) a pixel that has not been determined whether or not to constitute a character / line diagram is a “gorge” (minimum portion is linear in the luminance curved surface f (x, y)). Or a region that is continuous in the form of a band) or “recess” (region where the local minimum part is localized), a binary that reproduces or improves the character / diagram, assuming that the pixel constitutes the character / diagram An image can be generated.

The luminance gradient amount detection unit 207 can detect the luminance gradient in each pixel of the high resolution multilevel image generated by the original high resolution multilevel image generation unit 202.
The first contour correcting unit 2081 reproduces or improves the character / line diagram created by the topographic feature incorporating unit 206 based on the luminance gradient amount of each pixel detected by the luminance gradient amount detecting unit 207. The contour of the diagram can be modified.
The character / line diagram recognition unit 209 performs character / line diagram recognition of the image in which the outline of the character / line diagram is corrected by the first contour correction unit 2081. That is, the pattern recognition unit 2091 performs character / line diagram recognition with reference to the character / line diagram pattern stored in the pattern storage unit 2092.
The image output means 30 can output to a display device, a printing device, or an external device connected via a network.

<< Second Embodiment >>
FIG. 3 shows the configuration of the image recognition device 2B of the second embodiment.
In the second embodiment, the image recognition apparatus 2B includes a second contour correction unit 2082 immediately after the first contour correction unit 2081 in addition to the components of the image recognition device 2A. The second contour correcting means 2082 can be realized by a second contour correcting program.

  The second contour correcting unit 2082 clusters the pixel columns constituting the contour of the character / line diagram in the binary image corrected by the first contour correcting unit 2081 based on the tangent direction, and contours each cluster. By smoothing and / or sharpening the corners, it is possible to correct the outline of the character / line diagram in the binary image.

<< Third Embodiment >>
FIG. 4 shows the configuration of an image recognition device 2C according to the third embodiment.
In the third embodiment, the image recognition device 2C includes an original image storage unit 210 and a color restoration unit 211 in addition to the components of the image recognition device 2B. The original image storage unit 210 includes a storage device as hardware and an original image storage program, and the color restoration unit 211 can be realized by a color restoration program.

  The original image storage unit 210 stores the original image acquired by the original image acquisition unit 201, and the color restoration unit 211 specifies the color of the character / line diagram of the original image and uses this color as the character / line diagram. It is given to the character / line diagram recognized by the recognition means 209. Usually, the color assigned to the character / diagram recognized by the character / diagram recognition unit 209 is approximated to the color of the character / diagram of the original image.

<< 4th Embodiment >>
FIG. 5 shows the configuration of the image recognition device 2D of the fourth embodiment.
In the fourth embodiment, the image recognition device 2D, like the image recognition device 2A, is an original image acquisition unit 201, an original high-resolution multivalued image generation unit 202, a basic image generation unit 203, a luminance curved surface generation unit 204, a topographic feature. An extraction unit 205, a topographic feature incorporation unit 206, a brightness gradient amount detection unit 207, a first contour correction unit 2081, and a character / line diagram recognition unit 209 are provided.
In the present embodiment, the luminance curved surface generation unit 204 generates a luminance curved surface f (x, y) from the multivalued image generated by the original high resolution multilevel image generation unit 202.

<< 5th Embodiment >>
FIG. 6 shows the configuration of an image recognition device 2E according to the fifth embodiment.
In the fifth embodiment, the image recognition device 2E includes a second contour correction unit 2082 immediately after the first contour correction unit 2081 in addition to the components of the image recognition device 2D.

  Similar to the second embodiment, the second contour correcting unit 2082 clusters the pixel columns constituting the contour of the character / line diagram in the binary image corrected by the first contour correcting unit 2081 based on the tangent direction. Thus, the contour of the character / line diagram in the binary image can be corrected by smoothing the contour and / or sharpening the corners of each cluster.

<< 6th Embodiment >>
FIG. 7 shows a configuration of an image recognition device 2F according to the sixth embodiment.
In the sixth embodiment, the image recognition device 2F includes an original image storage unit 210 and a color restoration unit 211 in addition to the components of the image recognition device 2E.

  Similar to the third embodiment, the original image storage unit 210 stores the original image acquired by the original image acquisition unit 201, and the color restoration unit 211 specifies the color of the character / line diagram of the original image, This color is assigned to the character / line diagram recognized by the character / line diagram recognition unit 209. Usually, the color assigned to the character / diagram recognized by the character / diagram recognition unit 209 is approximated to the color of the character / diagram of the original image.

Hereinafter, an embodiment of the present invention will be described with reference to the flowchart of FIG.
<< Generation of basic image by interpolation and local statistics >>
The basic image is generated by the following procedure.
First, the original image acquisition unit 201 acquires an original image (multi-level bitmap image) (S101), and if the acquired image (multi-level bitmap image) is a grayscale image ("NO" in S102). ), And this is used as the processing target image as it is. If the acquired image is a color image (“YES” in S102), the original image is converted into a grayscale image IO (S103).

Next, a high-resolution multilevel image IH is generated by interpolation (S104). There are various methods for interpolation. Here, bilinear interpolation, which is easy to calculate, is used, and the result is subjected to smoothing (for example, 3 × 3 linear filter) processing. Further, for use in the first contour correction described later, a luminance gradient amount GH is calculated for each pixel on the high-resolution multilevel image IH, and the calculation result is stored in a predetermined memory (S105). .
Further, a basic image FH is generated from the high-resolution multilevel image IH based on the local statistics (S106).

  FIG. 10B shows the result of binarization of the original image shown in FIG. 10A using the Niblack binarization technique. As is clear from FIG. 10B, the horizontal stroke is lost as it is. For this reason, as a basic image FH, a binarization technique by Niblack (W. Niblack, An introduction to image processing, pp. 115-116, Englewood Cliffs, NJ: Prentice Hall, 1986) is expanded.

That is, the basic image FH takes, for example, one of 1, 0, and -1.
FH (x, y) = 1: “ON” (foreground or character stroke),
FH (x, y) = 0: “OFF” (background),
FH (x, y) = − 1: “TBD” (may be “ON”, later determined by topographic features),
An image taking one of the three values is generated.

FIGS. 11A and 11B (B is an enlarged view of the portion W of the image of FIG. 11A) show examples of the basic image FH. Here, “ON”, “OFF”, and “TBD” of FH (x, y) are represented by black, white, and gray, respectively.
Specifically, for the high-resolution multilevel image IH (x, y), calculation is performed within a window (processing region: for example, 3 × 3 pixels, 5 × 5 pixels, etc.) centered on the pixel to be processed. Based on the average luminance μ (x, y) and standard deviation σ (x, y), the basic image FH (x, y) is set according to the following rules.

(1) If IH (x, y) ≦ μ (x, y) + k0σ (x, y) (where k0 is a default parameter), FH (x, y) = 1 (set to “ON”) To do.
(2) μ (x, y) + k0σ (x, y) <IH (x, y) <μ (x, y) + k1σ (x, y) (where k0 <k1) and (x, y ) In the vicinity of FH (x, y) is “ON”, FH (x, y) = − 1 (set to “TBD”).
(3) In other cases, FH (x, y) = 0 (set to “OFF”).

“TBD” indicates that when the density of strokes is high, such as kanji, it is not possible to determine whether the color is white or black without examining topographic features.
Actually, the luminance varies greatly depending on the sampling position even in the same stroke. 12B and 12C are luminance profiles along the scan lines A and B in the vertical direction of the grayscale original image Io shown in FIG. As shown in FIGS. 12 (B) and 12 (C), the tendency is remarkable particularly in the places where the horizontal strokes are dense. Therefore, in the case of (2) (FH (x, y) = − 1 (“TBD”)), the processing area (window) for determining FH (x, y) from IH (x, y) is determined. The size may be within t pixels in the upper, lower, left, and right sides (t is a parameter determined according to the resolution enlargement ratio).

《Completing missing strokes by topographic features》
A luminance curved surface z = f (x, y) is generated from the grayscale image (multi-valued image) IO to be processed (S107), and the topographic features (ridges) on the luminance curved surface z = f (x, y) are generated. , Gorge, summit, depression, hillside) are extracted (S108).
Here, white is defined as having a larger z value than black. An important feature for stroke interpolation is that the z value is locally small, that is, the gorge (f (x, y) is minimal in one direction) and the depression (f (x, y) is minimal in all directions). ).

  The specific calculation method of these features is described in “Direct gray-scale extraction of features for character recognition” by Wang and Pavlidis (“L. Wang and T. Pavlidis, Direct gray-scale extraction of features for character recognition”). , IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 15, no. 10, pp. 1053-1067, 1993.) or Seong-Whan Lee and Young Joon Kim Using the technology of “Direct Extraction of Topographic Features for Gray Scale Character Recognition: IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 17, No. 7, pp.724-729, July 1995.” be able to.

For example, the topographic features as shown in FIG. 9 can be obtained from the gray scale image (multi-value image) IO. Next, using the topographic features thus obtained, the stroke is complemented to create a term resolution corrected image JH (S109).
For the pixel having the value “TBD” in the basic image FH (x, y), if the corresponding pixel in the grayscale image (multi-valued image) IO is “the canyon” or “the depression”, the basic image FH (x , Y) is reset to “ON”, otherwise “OFF”. That is, the feature obtained from the luminance curved surface z = f (x, y) is prioritized over the image obtained from the basic image FH (x, y).

The binary image thus obtained is shown in FIG. Compared with the image (see FIG. 10B) before incorporating the topographic features into the basic image FH (see FIG. 10B), the horizontal stroke of the portion indicated by C31 (“month”) in the character C3 (“still”), and the character C2 It can be seen that the horizontal stroke ("base") has been restored.
However, in the image of FIG. 13, the stroke of the character is sufficiently restored, but since the outline is not smooth, the quality of the character image is clearly poor.

<First contour correction>
As already described, it is necessary that the curvature is small (the radius of curvature is large) as a reference for a smooth contour line. Further, it is assumed that the contour line passes through a pixel having a maximum luminance gradient amount (that is, a boundary between the white area and the black area).

  Therefore, the contour line is corrected by using the above-described Active Contour or Snake algorithm, which was conceived at the end of the 1980s and is now a standard method for image processing. In this algorithm described above, an initial curve with the arc length s as a parameter is set to v (s) = (x (s), y (s)), and v (s) is set so that the next quantity is minimized. It corrects (S110).

E = ∫ (α (s) Econt + β (s) Ecurv + λ (s) Eimage) ds
Here, Econt (≧ 0) is a term for preventing the curve from shrinking and the point sequence is arranged at equal intervals, and Ecurv (≧ 0) is a term for reducing the curvature of the point sequence. , Image (≦ 0) is a term for increasing the gradient amount on the image IH, and can be expressed as −GH (v (s)) using the luminance gradient amount GH calculated first. FIG. 14 shows an example of the luminance gradient amount GH. In FIG. 14, the higher the luminance gradient amount GH, the higher the density is displayed. As shown in FIG. 14, it can be seen that the luminance gradient amount of the outline portion of the character is large.

The three parameters α, β, and γ can be set to different values for each point, but here, fixed values are used (α = β = γ = 1.0).
The contour is corrected by using the Active Contour algorithm for each contour line (boundary of “ON” pixels) obtained from the basic image FH. After all the contour lines are corrected, a binary image is generated from the contour lines by vector-raster conversion.

  FIG. 15 shows an image obtained by superimposing the luminance gradient amount GH shown in FIG. 14 and the contour of the image shown in FIG. FIG. 16 shows the result of contour correction. As is apparent from a comparison between the image of FIG. 16 and the image of FIG. 13, it can be seen that the smoothness of the contour line and the uniformity of the thickness of the stroke are improved.

《Second contour correction》
Active Contour's algorithm significantly improves the quality of character images, but horizontal and vertical line shakiness and dullness at the intersection of vertical and horizontal lines are observed.

These are points of great concern for human perception. Therefore, the contour can be shaped so that it looks beautiful to humans. In particular, the shakiness of the horizontal and vertical lines and the dullness of the intersection of the vertical and horizontal lines are corrected (S111).
Now, the closed contour to be processed is represented by a point sequence P = (p (0), p (1),..., P (n−1)). However, the subscripts in parentheses for p in the point sequence P are given for convenience. That is, considering that the subscript value i of the point p (i) is generally not 0 to n-1, if i> n, then p (i) = p (i−n), i < If 0, p (i) = p (i + n)).

In the present embodiment, this point sequence P is clustered in the tangential direction. With the tangential direction of the point p (i) as θ (i), a label L (i) is given to the point p (i) as follows.
(Jπ / 2) −δ ≦ θ ≦ (jπ / 2) + δ
(Where δ is a parameter sufficiently smaller than π / 2, j = 0, 1, 2, 3)
Then, L (i) = j. That is, if the tangential direction of the point p (i) is nearly 0 °, 90 °, 180 °, 270 °, the value of L (i) = 0, 1, 2, 3 is given to the point p (i). .

In other cases, L (i) = − 1.
That is, a point p having a tangent close to the horizontal direction has label 0 or 2, a point having a tangent close to the vertical direction has label 1 or 3, and the other points have label -1.
In this way, a series of levels (L (0), L (1),..., L (n−1)) is obtained from the point sequence P. By collecting a series of points having the same label as one cluster, the point sequence P can be clustered as shown in FIG.

For the jth cluster, let kj be the index of the starting point of that cluster, mj> 0 be the number of points belonging to that cluster,
Cj = (p (kj), p (kj + 1),..., P (kj + mj-1))
Then,
Between the (j-1) th and (j + 1) th clusters,
kj = kj-1 + mj-1, kj + 1 = kj + mj
There is a relationship
L (kj−1 + mj−1) ≠ L (kj) = L (kj + 1)
= ... = L (kj + mj-1) ≠ L (kj + 1)
Satisfies the property.
For the cluster Cj, the shakiness of the horizontal and vertical lines is corrected. If L (kj) = 0 or 2, that is, a cluster having a tangent line close to the horizontal direction, the y coordinate Y (kj + i) (i = 0, 1,..., Mj−1) of each point of Cj The mode M is obtained from the distribution of.

If | Y (kj + i) −M | ≦ 1, Y (kj + i) ← M is set.
When L (kj) = 1 or 3, the same processing is performed for the x coordinate.
Next, the dullness at the intersection of the vertical and horizontal lines is shaped. A cluster Cj of points whose tangent directions are not horizontal / vertical is surrounded by a cluster of points having tangential directions perpendicular to each other, that is, L (kj−1) ≧ 0, L (kj) <0, L ( kj + 1)> 0 and L (kj−1) ≠ (kj + 1).

If the number of points belonging to the cluster Cj is sufficiently small, as shown in FIG. 18B, the points in the cluster Cj are formed by extending the clusters Cj-1 and Cj + 1 to form a right-angled corner. To correct.
After all the contour lines are corrected, a binary image is generated from the contour lines by vector-raster conversion. The above processing flow is shown in FIG.

FIG. 19A shows a state before the kanji character “times” (indicated by reference numeral C5) is subjected to these shaping processes, and FIG. 19B shows a state after the kneading process. In FIGS. 19A and 19B, black pixels a, c, e, g, and i are pixels that do not constitute a cluster (pixels whose tangent lines are out of 0 °, 90 °, 180 °, and 270 °). , B, d, f, and h indicate pixels constituting the cluster (pixels whose tangent lines are close to 0 °, 90 °, 180 °, and 270 °).
It can be seen that the unevenness of the contour line (pixel row) before the processing shown in FIG. 19A and the dullness of the corner portion (indicated by the symbol CNR) are alleviated in the contour line shown in FIG.

  FIG. 20B shows the result of creating the 200 dpi original image shown in FIG. As can be seen from FIGS. 20A and 20B, according to this embodiment, even if the multi-tone original image is not clear, a binary bitmap image having excellent reproducibility is reproduced. Note that the recognition accuracy of 97.2% was obtained with the method of Patent Document 1 for a color image input at 200 dpi, but the recognition accuracy of 99.1% was obtained with the above method.

《Image recognition》
The image after the above processing is recognized by the image recognition program GRP (S112).

It is a figure which shows the example of 1 structure of the image recognition apparatus of this invention. It is a functional block diagram which shows the flow of the process by each means in 1st Embodiment. It is a functional block diagram which shows the flow of the process by each means in 2nd Embodiment. It is a functional block diagram which shows the flow of the process by each means in 3rd Embodiment. It is a functional block diagram which shows the flow of the process by each means in 4th Embodiment. It is a functional block diagram which shows the flow of the process by each means in 5th Embodiment. It is a functional block diagram which shows the flow of the process by each means in 6th Embodiment. It is a flowchart which shows the Example of this invention. It is a figure which shows the topographic feature extracted from xy plane of a luminance curved surface. (A) is a figure which shows the result which binarized the original image and (B) the image of (A) using the binarization technique by Niblack. (A) is a figure which shows the example of a basic image, (B) is the elements on larger scale of the image of (A). (A) is a gray scale original image, (B) is a luminance profile along the vertical scan line A of the gray scale original image of (A), and (C) is a luminance profile along the scan line B. is there. It is a figure which shows the binary image which took in the topographical feature of the luminance curved surface z = f (x, y). It is a figure which shows the example of the brightness | luminance gradient in each pixel of a high-resolution multi-value image. It is a figure which shows the image which overlap | superposed the brightness | luminance gradient amount shown in FIG. 13, and the image shown in FIG. It is a figure which shows the result of having performed the process by the 2nd outline correction means about the image shown in FIG. It is a figure which shows the processing flow when producing | generating a binary image from an outline by vector-raster conversion, after correcting about all the outlines. (A) It is a figure which shows the example which divided the outline pixel row into clusters, (B) is a figure which shows the example which corrected the unevenness | corrugation per cluster. . (A) is a figure which shows the state before performing the shaping process about a Chinese character, (B) is a figure which shows the state after performing the shaping process. FIG. 6A is a diagram illustrating a 200 dpi original image, and FIG. 5B is a diagram illustrating a result created at a resolution of 4 times.

Explanation of symbols

2A-2J Image recognition apparatus 30 Image output means 100 Communication line 110 PC (personal computer)
111 CPU
DESCRIPTION OF SYMBOLS 112 Memory 113 Hard disk apparatus 114 Removable disk apparatus 115 Display 116 Printer 117 Keyboard 118 Network interface 119 Bus 201 Original image acquisition means 202 High resolution multi-value image generation means 203 Basic image generation means 204 Luminance curved surface generation means 205 Topographic feature extraction means 205 206 Topographical feature incorporation means 207 Luminance gradient amount detection means 209 Character / line diagram recognition means 210 Original image storage means 211 Color restoration means 1121 ROM
1122 RAM
2011 color / grayscale conversion means 2081 first contour correction means 2082 second contour correction means 2091 pattern recognition means 2092 pattern storage means

Claims (22)

Original image acquisition means for acquiring an original image of a multi-valued bitmap;
From the original image acquired by the original image acquisition means, a high-resolution multi-value image generation means for generating a multi-value image having a higher resolution than the original image;
A luminance curved surface generating means for generating a curved surface in which the xy coordinates are pixel coordinates and the z coordinate is luminance from the original image acquired by the original image acquiring means;
Topographic feature extraction means for extracting topographic features of the luminance curved surface generated by the luminance curved surface generation means;
The character / line diagram is reproduced by incorporating the topographic feature extracted by the topographic feature extracting unit into the outline of the character / line diagram included in the multi-level image generated by the high-resolution multi-level image generating unit. Or a topographic feature embedding means for generating an improved binary image;
Contour correcting means for correcting the contour of a character / line diagram included in the binary image generated by the topographic feature incorporating means;
A character / line diagram recognition unit for recognizing the character / line diagram as a character code / vector diagram of the image in which the contour of the character / line diagram is corrected by the contour correction unit;
An image recognition apparatus comprising: The topographical features include a “valley or depression” having lower luminance than the surroundings when the luminance curved surface is made to correspond to actual terrain, “ridge or mountain peak”, “valley or depression” having higher luminance than the surroundings, and “ The image recognition apparatus according to claim 1, wherein the image recognition apparatus is a “mountain hill or a buttocks” located between the “ridge or the mountain top”. Furthermore, a color restoration means is provided,
The color restoration unit restores the character / line diagram recognized by the character / line diagram recognition unit or the background color by approximating the color of the original image acquired by the original image acquisition unit. The image recognition apparatus according to claim 1 or 2. 4. The original image acquisition means comprises color / grayscale conversion means for converting the original image into a grayscale multilevel image when the original image is a color image. The image recognition apparatus according to any one of the above. The topographic feature incorporation means includes a character / line in the original image as a part of a character / line diagram including a region where the minimal part of the luminance curved surface is continuous in a linear or belt shape and a region where the minimal part is localized. The image recognition apparatus according to claim 1, wherein the binary image in which a figure is reproduced or improved is generated. From the high-resolution multi-value image generated by the high-resolution multi-value image generating means, based on statistical information referring to surrounding pixels of each pixel, (a) pixels constituting a character / line diagram, and (b) characters / Basic image generation means for generating a basic image (an image in which topographic features are incorporated) composed of pixels that do not constitute a diagram and (c) pixels that are not determined whether or not to constitute a character / line diagram With
6. The topographic feature embedding unit generates a binary image in which the character / line diagram is reproduced or improved based on the basic image generated by the basic image generating unit. The image recognition apparatus described in 1. The topographic feature incorporation means is characterized in that (c) a pixel for which whether or not to constitute a character / line diagram is determined is a region in which a minimal portion of the luminance curved surface is continuous in a linear or belt shape or the minimal portion 7. The binary image obtained by reproducing or improving the character / diagram is generated on the assumption that the pixel constitutes the character / diagram when the pixel is included in the localized region. Image recognition device. A luminance gradient amount detecting means for detecting a luminance gradient amount in each pixel of the high resolution multilevel image generated by the high resolution multilevel image generating unit;
The contour correcting unit reproduces or improves the character / line diagram created by the topographic feature incorporation unit based on the luminance gradient amount of each pixel detected by the luminance gradient amount detecting unit. First contour correcting means for correcting
The image recognition apparatus according to claim 1, further comprising: The contour correcting means refers to a curvature or direction change of a pixel row that forms an outline of the character / diagram in the image corrected by the first contour correcting means, and further changes the character / diagram in the image. 9. The image recognition apparatus according to claim 8, further comprising second contour correcting means for correcting the contour. The second contour correcting unit clusters the pixel columns constituting the contour of the character / diagram in the image corrected by the first contour correcting unit based on a tangent direction, and the contour for each cluster. The image recognition apparatus according to claim 9, wherein the contour of the character / line diagram in the image is corrected by smoothing and / or sharpening the corners. Original image acquisition means for acquiring an original image of a multi-valued bitmap;
From the original image acquired by the original image acquisition means, a high-resolution multi-value image generation means for generating a multi-value image having a higher resolution than the original image;
A luminance curved surface generating means for generating a curved surface in which the xy coordinates are pixel coordinates and the z coordinate is luminance from the multi-value image generated by the high-resolution multi-value image generating means;
Topographic feature extraction means for extracting topographic features of the luminance curved surface generated by the luminance curved surface generation means;
Topographic feature incorporation means for creating a binary image that reproduces or improves a character / line diagram in the original image by incorporating the topographic feature extracted by the topographic feature extraction means;
Contour correcting means for correcting the contour of a character / line diagram included in the binary image generated by the topographic feature incorporating means;
Contour correcting means for correcting the contour of a character / line diagram included in the binary image generated by the topographic feature incorporating means;
A character / line diagram recognition unit for recognizing the character / line diagram of the image whose contour has been corrected by the contour correction unit as a character code / vector diagram;
An image recognition apparatus comprising: The topographical features include a “valley or depression” having lower luminance than the surroundings when the luminance curved surface is made to correspond to actual terrain, “ridge or mountain peak”, “valley or depression” having higher luminance than the surroundings, and “ The image recognition apparatus according to claim 11, wherein the image recognition apparatus is a “mountain hill or a buttocks” located between the “ridge or the mountain top”. Further, the character / diagram or the background color of the image output by the character / diagram recognition means is generated by approximating the color in the original image acquired by the original image acquisition means. The image recognition apparatus according to claim 11 or 12. 14. The original image acquisition means comprises color / grayscale conversion means for converting the original image into a grayscale multilevel image when the original image is a color image. The image recognition apparatus according to any one of the above. The topographic feature incorporation means includes a character / line in the original image as a part of a character / line diagram including a region where the minimal part of the luminance curved surface is continuous in a linear or belt shape and a region where the minimal part is localized. 15. The image recognition apparatus according to claim 11, wherein the binary image is generated by reproducing or improving a character / line diagram obtained by reproducing or improving a figure. From the high-resolution multi-value image generated by the high-resolution multi-value image generating means, based on statistical information referring to surrounding pixels of each pixel, (a) pixels constituting a character / line diagram, and (b) characters / Comprising basic image generating means for generating a basic image consisting of pixels that do not constitute a diagram and (c) pixels that are not determined whether or not to constitute a character / diagram;
16. The topographic feature incorporation means generates a binary image in which the character / line diagram is reproduced or improved based on the basic image generated by the basic image generation means. The image recognition apparatus described in 1. The topographic feature incorporation means is characterized in that (c) a pixel for which whether or not to constitute a character / line diagram is determined is a region in which a minimal portion of the luminance curved surface is continuous in a linear or belt shape or the minimal portion 17. The binary image obtained by reproducing or improving a character / diagram is generated assuming that the pixel constitutes a character / diagram when the pixel is included in a localized region. Image recognition device. A luminance gradient amount detecting means for detecting a luminance gradient amount in each pixel of the high resolution multilevel image generated by the high resolution multilevel image generating unit;
The contour correcting unit reproduces or improves the character / line diagram created by the topographic feature incorporation unit based on the luminance gradient amount of each pixel detected by the luminance gradient amount detecting unit. First contour correcting means for correcting
The image recognition apparatus according to claim 11, further comprising: The contour correcting means refers to a curvature or direction change of a pixel row that forms an outline of the character / diagram in the image corrected by the first contour correcting means, and further changes the character / diagram in the image. The image recognition apparatus according to claim 18, further comprising second contour correcting means for correcting the contour. The second contour correcting unit clusters the pixel columns constituting the contour of the character / diagram in the image corrected by the first contour correcting unit based on a tangent direction, and the contour for each cluster. 20. The image recognition apparatus according to claim 19, wherein the contour of the character / line diagram in the image is corrected by smoothing and / or sharpening the corners. An image recognition program for causing a computer to function as each means according to any one of claims 1 to 20. A computer-readable recording medium recording a program for causing a computer to be executed as each means according to any one of claims 1 to 20.

Images