JP5028911B2 - Character string recognition program, method and apparatus - Google Patents

Description

  The present invention generally relates to a program, method and apparatus for character recognition, and more particularly to a character string recognition program, method and apparatus for a deteriorated character string.

With the widespread use of digital cameras and digital video cameras for capturing document images, the recognition of degraded character strings is gaining more and more attention. The recognition of a deteriorated character string includes single character recognition and character extraction from a character string, and these two parts are originally combined.

  Regarding character segmentation from a character string, a segmentation method based on recognition is the most widely used method. FIG. 1 shows the principle of a conventional clipping method based on recognition. The input image is first binarized and then the pixel connected components of the binarized image are analyzed to find the stroke of the character (top row of FIG. 1). Refer to the following Non-Patent Document 1 for the analysis algorithm of the pixel connected component of the image. All pixel connected components are considered basic segment characters (middle row in FIG. 1). The combination of pixel connected components is regarded as a composite segment character (bottom line in FIG. 1). Subsequently, character recognition is performed on all basic segment characters as well as composite segment characters, thereby providing a recognition distance. All character strings can be separated into multiple cutout paths consisting of different basic segment characters and combined composite segment characters, and the recognition distance of each cutout path is the recognition of the basic segment characters and composite segment characters that compose it The total distance. An accurate cutout result of the character string is obtained by selecting a cutout path having the minimum total recognition distance. While achieving segmentation, the recognition result for each basic segment character and composite segment character is also the final recognition result for the character.

  FIG. 1 shows the principle of a conventional clipping method based on recognition.

  As shown in FIG. 1, the cut-out path made up of “C”, “Li”, and “O” has a minimum recognition distance 72. Therefore, these are output as the final cutout and recognition results.

  From the above principle, it can be seen that the recognition distance is very important not only for the recognition result but also for accurate extraction. For example, in FIG. 1, the minimum recognition distance of “C” is 21, and the recognition distances of the two left and right strokes of the character are 19 and 26, respectively. If the sum of the recognition distances of these two strokes is less than 21, even if the first recognition result of “c” is accurate, it is still erroneously segmented into left stroke 1 and right stroke 2 in FIG. become.

  Numerous papers and patents related to character segmentation from character strings such as the following patent documents 1 and 2 and non-patent documents 1 to 3 have been published so far.

  Many of these papers and patents aim to process contact characters, and many of the objects to be processed are binarized images. For degraded character string images, conventional binarization methods often cause significant blurred strokes (missing stroke pixel points) and stroke contact, making it impossible to achieve the desired recognition effect It is.

  The dual eigenspace-based method is very effective in recognizing degraded characters, and this method extracts character features directly from grayscale character images. FIG. 2 is a flowchart illustrating character recognition using a dual eigenspace-based method. The input is a normalized character image. The feature of the character image is first extracted by the first dictionary (dictionary 1 in FIG. 2). The character images are then roughly classified into M category candidates by the second dictionary (dictionary 2 in FIG. 2). A third dictionary (dictionary 3 in FIG. 2) is then used to finally classify the input character features into one of M category candidates. Finally, the recognized character code and the recognition distance are output.

  The dual eigenspace based method is more stable against image noise due to defects because it avoids the binarization process by extracting features directly from the grayscale image. However, there are some problems when the double eigenspace based method is directly applied to the recognition based clipping method.

  As shown in FIG. 3, the image in the first row is a character string image. The second row is the binary result, and this binary image is used for rough segmentation. The circumscribed rectangle as shown is the result of the rough cut. The third row is a normalized grayscale image of basic segment characters. Shown below each segment image is the recognized character and its recognition distance. The fourth row is a normalized grayscale character image of the composite segment characters “year” and “open” after normalization, and the corresponding recognition result and recognition distance. When the conventional cut-out method based on recognition is used, “open” in the second row is separated into four segments, so “open” is not recognized correctly, and four segments are recognized. The total distance is 5.39 + 61.01 + 45.69 + 20.37 = 132.46. The recognition distance of “open” is 409.71, which is larger than the sum of the recognition distances of the four segments, so that the entire character string is recognized as “I! II” once a year.

US Pat. No. 6,327,385 U.S. Patent No. 5,691,069 US Pat. No. 5,172,422 R. C. Page 435 of "digital Image Processing second edition" by Gonzalez, edited by Quiqi RUAN, Yuzhi RUAN et al. Y. Lu, “Machine Printed Character Segmentation—An Overview”, Pattern Recognition, Vol. 28, No. 1, pp. 67-80, January 1995. S. W. Lee, D.C. J. et al. Lee, H.C. S. Park, “A New Methodology for Gray-Scale Character Segmentation and Recognition”, IEEE transaction on pattern analysis, 10th, 10th, 10th, 10th, 10th, 10th, 10th, 10th, 10th, 10th.

  An object of the present invention is to provide a character string recognition apparatus and method for a deteriorated character string that uses a better feature to generate a more reasonable recognition distance to solve the problems associated with clipping using a double eigenspace. Is to provide.

In accordance with one aspect of the present invention, feature extraction means for extracting features from a normalized image input using a first dictionary composed of a transformation matrix and an average value of each normalized image, and the extracted A coarse classification means for selecting a certain number of character category candidates by comparing the features and the features stored in the second dictionary; and a third classification means for storing a conversion matrix and an average feature vector for each character category . Using the dictionary and the predetermined number of selected character category candidates, feature reconstructing means for constructing a certain number of reconstructed features, features extracted by the feature extracting means and the reconstructed Using a detailed classification means for recognizing and outputting a character code recognized according to the feature, the first dictionary, and the reconstructed feature generated by the feature reconstructing means, Reconfiguration An image reconstruction means for constructing an image, a normalized image that is the input, a recognition distance calculating means for calculating and outputting a recognition distance in accordance with the reconstruction image generated by the image reconstruction means, said recognition distance A character string recognition unit that performs character segmentation based on the recognition distance calculated by the calculation unit and recognizes the character code calculated by the detailed classification unit corresponding to the character segmentation result as a final recognition character code. A character string recognition device for a character string is provided.

In accordance with another aspect of the present invention, extracting features from a normalized image input using a first dictionary composed of a transformation matrix and an average value of each normalized image, and the extracted features Selecting a certain number of character category candidates by comparing the features stored in the second dictionary; a third dictionary storing a conversion matrix and an average feature vector for each character category ; and Using a fixed number of selected character category candidates, constructing a fixed number of reconstructed features, and recognizing recognized character codes according to the extracted features and the reconstructed features And using the first dictionary and the reconstructed features to construct a fixed number of reconstructed images, the input normalized image and the reconstructed To the image Calculating and outputting a recognition distance I performs character extraction based on the recognition distances said calculated recognize a character code corresponding to said character extraction result of the recognized character code as the last recognized character codes and a step of, string recognition method for degradation strings are provided.

  It is possible in the present invention to recognize and output the final recognition character code according to the extracted feature and the reconstructed feature, and to calculate and output the recognition distance according to the input normalized image and the reconstructed image. For this reason, the present invention makes it possible to correctly segment the characters of the degraded character string by using a better feature and generating a recognition distance more suitable for clipping.

  According to the present invention, the final recognition character code is recognized and output according to the extracted feature and the reconstructed feature, and the recognition distance is calculated and output according to the input normalized image and the reconstructed image. As a result, it is possible to correctly segment the characters of the degraded character string image.

  Embodiments of the present invention will be described below with reference to the accompanying drawings.

  FIG. 4 is a flowchart showing a character string recognition method used by the character string recognition device according to the embodiment of the present invention.

  As shown in FIG. 4, the character string recognition apparatus according to the embodiment of the present invention includes a feature extraction unit 402 that extracts features from a normalized image 401 input using the first dictionary 403, The coarse classification unit 404 that selects M character category candidates by comparing the features stored in the second dictionary 405 with the extracted features, and the third dictionary 407 and the M character category candidates. Extracted by the feature reconstruction unit 406 that reconstructs the M reconstruction features, the image reconstruction unit 408 that configures the M reconstruction images using the first dictionary 403, and the feature extraction unit 402 A detailed classification unit 409 that outputs a character code 411 recognized by comparing the difference between the reconstructed feature and the reconstructed feature, and a recognition distance calculation unit 410 that outputs a recognition distance 412.

  According to the flowchart shown in FIG. 4, for the input normalized character image 401, the feature extraction unit 402 uses the following expression to calculate the features of the normalized character image 401 input using the first dictionary 403. Extract according to (1):

Here, X = [x ₁ , x ₂ ,..., X _{w * h} ] ^T represents a normalized character image whose height and width are w and h, respectively. Equation (2) is the average value of all normalized character images. _{U = [u 1, u 2} , ..., u n] is _{u i = [u i1, u} i2, ..., u iw * h] is a transformation matrix for the ^T. The first dictionary 403 consists of U and Equation (3). The feature extraction method shown in equation (1) is called principal component analysis (PCA). For more information on PCA, see “Pattern Classification”, Second edition, by R.C. O. Duda, P.A. E. Hart and D.H. G. Stork, a Wiley-Interscience Publication, John Wiley & Sons, Inc. 2001.115-117, 568-569.

After the feature extraction, the extracted feature Y is compared with the features of each character category stored in the second dictionary 405 by the coarse classification unit 404. There are a number of algorithms for feature comparison, one of which is the Euclidean distance: D _i = | Y−Y _i | (where D _i is the Euclidean distance of the feature Y up to the feature of the i th character category Y _i. ). When the number of candidate character categories output from the coarse classification unit 404 is M, M character categories are selected as the coarse classification output in order from the smallest Euclidean distance for each segment character.

Next, the feature reconstruction unit 406 uses the third dictionary 407 to construct M reconstruction features corresponding to the M candidate categories. The third dictionary stores the transformation determinant (4) and average feature vector C _i for each character category. The i-th reconstruction feature equation (5) is obtained by equation (6).

  The detailed classification unit 409 in FIG. 4 calculates the distance between the original feature Y and the M reconstruction feature equations (5). Of the M character categories calculated for each segment character, a code corresponding to the character category having the minimum distance is output as the recognized character code 411 of the segment character.

  Unlike the conventional method using a double eigenspace as shown in FIG. 2, the recognition distance of the present invention is not the difference between the extracted feature Y and the reconstructed feature. In the present invention, the new image reconstruction unit 408 uses the first dictionary 403 to calculate M reconstructed image equations (9) according to the following equations (7) and (8). Propose.

  Equation (7) can be derived from Equation (1). Equation (8) is used to normalize the range of pixel values in the reconstructed image to 0-255, which matches the range of pixel values in the original image.

  The recognition distance calculation unit 410 in FIG. 4 calculates the distance between the original normalized character image 401 and the M reconstructed image equations (9). The minimum distance is regarded as the recognition distance 412 that is finally output. In this way, character segmentation is performed so that the recognition distance 412 is minimized, and among the recognized character codes 411 output from the detailed classification unit 409, the character code corresponding to the character segmentation is finally recognized. It becomes a character code.

  FIG. 5 shows the recognition distance obtained by the character string recognition method used in the character string recognition device according to the embodiment of the present invention. It can be seen that the recognition distance in FIG. 5 is more reasonable for clipping. The recognition distance of “open” is 104.78, while the total recognition distance of its four components is 494.02, which is much larger than 104.78. This character can therefore be correctly segmented and recognized.

  Although the exemplary characters used in this embodiment are Japanese characters, the method presented in the present invention is not limited to Japanese only. This is also applicable to other languages such as Chinese and Korean.

(Supplementary note 1) A character string recognition program for a deteriorated character string,
A feature extraction procedure for extracting features from the normalized image input using the first dictionary;
A coarse classification procedure for selecting a certain number of character category candidates by comparing the extracted features with features stored in a second dictionary;
A feature reconstruction procedure for constructing a fixed number of reconstructed features using a third dictionary and the fixed number of selected character categories;
In addition,
A detailed classification procedure for recognizing and outputting the character code recognized according to the feature extracted by the feature extraction procedure and the reconstructed feature;
An image reconstruction procedure for constructing a fixed number of reconstructed images using the first dictionary and the reconstruction features generated by the feature reconstruction procedure;
A recognition distance calculation procedure for calculating and outputting a recognition distance according to the input normalized image and the reconstructed image generated by the image reconstruction procedure;
A character string recognition program that causes a computer to execute.

(Supplementary note 2) Character extraction is performed based on the recognition distance calculated in the recognition distance calculation procedure, and the character code calculated in the detailed classification procedure corresponding to the character extraction result is calculated as a final recognition character code. The character string recognition program according to Supplementary Note 1, which is characterized.

(Supplementary Note 3) The detailed classification procedure compares the difference between the feature extracted by the feature extraction procedure and the reconstructed feature, and the character code corresponding to the reconstructed feature having the minimum difference is recognized. The character string recognition program according to appendix 1 or 2, characterized in that:

(Additional remark 4) The said character reconstruction procedure normalizes the range of the pixel value of the said reconstructed image to 0-255, The character string recognition program of Additional remark 1 or 2 characterized by the above-mentioned.

(Supplementary Note 5) The recognition distance calculation procedure calculates a distance between the input normalized image and the reconstructed image generated by the image reconstruction procedure, and outputs a minimum distance as a recognition distance. The character string recognition program according to Supplementary Note 1 or 2, which is a feature.

(Additional remark 6) The said 1st dictionary is comprised from the conversion matrix and average value of each normalized image, The character string recognition program as described in any one of Additional remark 1-5 characterized by the above-mentioned.

(Additional remark 7) The said 2nd dictionary memorize | stores the characteristic of each character category, The character string recognition program as described in any one of additional remarks 1-5 characterized by the above-mentioned.

(Supplementary note 8) The character string recognition program according to any one of Supplementary notes 1 to 5, wherein the third dictionary stores a conversion matrix and an average feature vector of each character category.

(Supplementary note 9) A character string recognition method for a bad text string character string,
Extracting features from the normalized image input using the first dictionary;
Selecting a certain number of character category candidates by comparing the extracted features with features stored in a second dictionary;
Configuring a fixed number of reconstructed features using a third dictionary and the fixed number of selected character categories;
Recognizing and outputting a character code recognized according to the extracted features and the reconstructed features;
Constructing a fixed number of reconstructed images using the first dictionary and the reconstructed features;
Calculating and outputting a recognition distance according to the input normalized image and the reconstructed image;
A character recognition sequence method comprising:

(Supplementary note 10) Character extraction is performed based on the recognition distance calculated in the recognition distance calculation step, and the character code calculated in the detailed classification step corresponding to the character extraction result is calculated as a final recognition character code. The character string recognition method according to appendix 9, which is a feature.

(Supplementary Note 11) The step of recognizing and outputting the recognized character code includes:
Comparing a difference between the extracted feature and the reconstructed feature, and outputting a character code corresponding to the reconstructed feature having a minimum difference as a recognized character code; The character string recognition method according to Supplementary Note 9 or 10.

(Supplementary note 12) The step of reconstructing the reconstructed image comprises:
11. The character string recognition method according to appendix 9 or 10, comprising a step of normalizing a range of pixel values of the reconstructed image to a range of 0 to 255.

(Supplementary Note 13) The step of calculating and outputting the recognition distance includes:
The character string recognition method according to appendix 9 or 10, further comprising a step of calculating a distance between the input normalized image and the reconstructed image and outputting a minimum distance as a recognition distance.

(Additional remark 14) The said 1st dictionary is comprised by the conversion matrix and average value of each normalized image, The character string recognition method as described in any one of additional marks 9-13 characterized by the above-mentioned.

(Supplementary note 15) The character string recognition method according to any one of supplementary notes 9 to 13, wherein the second dictionary stores characteristics of each character category.

(Supplementary note 16) The character string recognition method according to any one of supplementary notes 9 to 13, wherein the third dictionary stores a conversion matrix and an average value of each character category.

(Supplementary Note 17) A character string recognition device for a defective text string character string,
Feature extraction means for extracting features from the normalized image input using the first dictionary;
Coarse classification means for selecting a certain number of character category candidates by comparing the extracted features with features stored in a second dictionary;
Feature reconstructor means for constructing a fixed number of reconstructed features using a third dictionary and the fixed number of selected character categories;
In addition,
Fine recognition detailed classification means for recognizing and outputting the character code recognized according to the feature extracted by the feature extraction means and the reconstructed feature;
Image reconstructing means for constructing a certain number of reconstructed images using the first dictionary and the reconstructed features generated by the feature reconstructing means;
Recognition distance calculation means for calculating and outputting a recognition distance according to the input normalized image and the reconstructed image generated by the image reconstruction means;
A character string recognition apparatus comprising:

(Supplementary note 18) Character extraction is performed based on the recognition distance calculated by the recognition distance calculation means, and the character code calculated by the detailed classification means corresponding to the character extraction result is calculated as a final recognition character code. The character string recognition device according to Supplementary Note 17, which is a feature.

(Additional remark 19) The recognition detailed classification means compares the difference between the feature extracted by the feature extraction means and the reconstructed feature, and the character code corresponding to the reconstructed feature having the minimum difference is recognized. 19. The character string recognition device according to appendix 17 or 18, wherein the character string recognition device outputs the code.

(Supplementary note 20) The character string recognition device according to supplementary note 17 or 18, wherein the image reconstruction unit normalizes a range of pixel values of the reconstructed image to 0 to 255.

(Supplementary Note 21) The recognition distance calculation unit calculates a distance between the input normalized image and the reconstructed image generated by the image reconstruction unit, and outputs a minimum distance as a recognition distance. 19. The character string recognition device according to appendix 17 or 18, which is a feature.

(Additional remark 22) The said 1st dictionary is comprised from the conversion matrix and average value of each normalization image, The character string recognition apparatus as described in any one of additional remarks 17-21 characterized by the above-mentioned.

(Additional remark 23) The said 2nd dictionary memorize | stores the characteristic of each character category, The character string recognition apparatus as described in any one of Additional remarks 17-21 characterized by the above-mentioned.

(Additional remark 24) The said 3rd dictionary memorize | stores the conversion matrix and average feature vector of each character category, The character string recognition apparatus as described in any one of additional marks 17-21 characterized by the above-mentioned.

It is a figure which shows the principle of the conventional clipping method based on recognition. FIG. 6 is a flowchart illustrating character recognition using a dual eigenspace-based method. It is a figure which shows an example of the character recognition which uses double eigenspace. It is a flowchart which shows the character string recognition method used by the character string recognition apparatus according to embodiment of this invention. It is a figure which shows an example of the character string recognition method used by the character string recognition apparatus according to embodiment of this invention.

Explanation of symbols

1 Left stroke 2 Right stroke

Claims (7)

A character string recognition program for a deteriorated character string,
A feature extraction procedure for extracting features from a normalized image input using a first dictionary composed of a transformation matrix and an average value of each normalized image ;
A coarse classification procedure for selecting a certain number of character category candidates by comparing the extracted features with features stored in a second dictionary;
Feature reconstruction procedure for constructing a fixed number of reconstructed features using a third dictionary storing a transformation matrix and an average feature vector for each character category and the fixed number of selected character category candidates When,
In addition,
A detailed classification procedure for recognizing and outputting the character code recognized according to the feature extracted by the feature extraction procedure and the reconstructed feature;
An image reconstruction procedure for constructing a fixed number of reconstructed images using the first dictionary and the reconstruction features generated by the feature reconstruction procedure;
A recognition distance calculation procedure for calculating and outputting a recognition distance according to the input normalized image and the reconstructed image generated by the image reconstruction procedure;
A character string recognition procedure that performs character segmentation based on the recognition distance calculated in the recognition distance calculation procedure and recognizes the character code calculated in the detailed classification procedure corresponding to the character segmentation result as a final recognition character code. A character string recognition program which is executed by a computer. The detailed classification procedure compares the difference between the feature extracted by the feature extraction procedure and the reconstructed feature, and outputs a character code corresponding to the reconstructed feature having the minimum difference as a recognized character code. The character string recognition program according to claim 1 . The character string recognition program according to claim 1, wherein the image reconstruction procedure normalizes a range of pixel values of the reconstructed image to 0 to 255. The recognition distance calculation procedure calculates a distance between the input normalized image and the reconstructed image generated by the image reconstruction procedure, and outputs a minimum distance as a recognition distance. Item 12. A character string recognition program according to Item 1 . Said second dictionary character string recognition program according to any one of claims 1-4, characterized by storing the characteristics of each character category. A character recognition sequence method for a degraded character string,
Extracting features from a normalized image input using a first dictionary composed of a transformation matrix and an average value of each normalized image ;
Selecting a certain number of character category candidates by comparing the extracted features with features stored in a second dictionary;
Configuring a fixed number of reconstructed features using a third dictionary storing a transformation matrix and an average feature vector for each character category and the fixed number of selected character category candidates ;
Recognizing and outputting a character code recognized according to the extracted features and the reconstructed features;
Constructing a fixed number of reconstructed images using the first dictionary and the reconstructed features;
Calculating and outputting a recognition distance according to the input normalized image and the reconstructed image;
Performing character segmentation based on the calculated recognition distance, and recognizing a character code corresponding to the character segmentation result among the recognized character codes as a final recognition character code. Recognition method. A character string recognition device for a deteriorated character string,
A feature extraction means for extracting features from a normalized image input using a first dictionary composed of a transformation matrix and an average value of each normalized image ;
Coarse classification means for selecting a certain number of character category candidates by comparing the extracted features with features stored in a second dictionary;
Feature reconstructing means for constructing a fixed number of reconstructed features using a third dictionary storing a transformation matrix and an average feature vector for each character category and the fixed number of selected character category candidates When,
In addition,
Detailed classification means for recognizing and outputting the character code recognized according to the feature extracted by the feature extraction means and the reconstructed feature;
It said first dictionary, using said reconstructed feature generated by the feature reconstruction means, and an image reconstruction means for constructing a plurality of reconstructed images of a certain number,
A normalized image that is the input, a recognition distance calculating means for calculating and outputting a recognition distance in accordance with said reconstructed image generated by the image reconstruction means,
A character string recognition unit that performs character segmentation based on the recognition distance calculated by the recognition distance calculation unit and recognizes the character code calculated by the detailed classification unit corresponding to the character segmentation result as a final recognition character code. A character string recognition device comprising:

Images