JP4231375B2 - A pattern recognition apparatus, a pattern recognition method, a pattern recognition program, and a recording medium on which the pattern recognition program is recorded. - Google Patents

Description

  The present invention relates to a pattern recognition technique for accurately identifying a two-dimensional pattern included in a natural image.

  In recent years, mobile phones have been equipped with a camera function, which makes it easy to obtain and store photo data at any time. As a means for further utilizing this function, a service of extracting character information from photograph data and providing data related to the information can be considered. At this time, one of the typical problems of pattern recognition is character recognition. Character recognition has long been studied for the purpose of print type OCR, handwritten character input interface, and the like. In such an application, there is no problem even if the input data is limited to character data. However, in the service as described above, there is a possibility that various textures, patterns and the like other than characters may be input. Therefore, it is necessary to improve the recognition method so as to eliminate the error of recognizing a non-character as a character.

The processing procedure for character recognition is roughly divided into three steps: “pre-processing unit”, “feature extraction unit”, and “identification unit”. Among them, the feature extraction unit is an important process that affects the recognition performance. As a feature extraction method in conventional character recognition, there is a “local direction histogram feature” (for example, see Non-Patent Document 1). This is by dividing the image data into several local blocks, calculating the edge direction component in each block, and creating the frequency distribution of edges quantized in four directions to calculate the feature amount It is a method to do. It has been confirmed that accurate recognition is possible mainly in handwritten character recognition.
Wakabayashi Satoshi, Tsuruoka Shinji, Kimura Fumitaka, Miyake Koji, “Higher accuracy of handwritten digit recognition by increasing the number of dimensions of features”, Science theory D-II, Vo1. J77-D-II, no. 10, pp. 2046-2053, 1994

  However, for example, when a complex texture is input as data other than characters, there is a problem that it is erroneously recognized as a complex character. An example of erroneous recognition is shown in FIG. This is considered to be due to the fact that the relative position information of the edges in the block is crushed because the histogram of the direction component is taken in each block. That is, in order to realize character recognition in a natural image, it can be said that a feature extraction method that clearly shows the difference between characters and those that do not is essential. Up to this point, the discussion has been limited to character recognition, but this is a common problem in the recognition of all patterns and objects in natural images.

  The present invention has been made in view of the above-described problems, and has recorded a pattern recognition device, a pattern recognition method, a pattern recognition program, and a pattern recognition program that can solve the above problems and enable more accurate recognition. An object is to provide a recording medium.

  In order to achieve the above object, the present invention introduces a new feature extraction method. Most patterns that are recognized by humans, such as characters, have long edge components connected to each other. FIG. 11 shows an edge extracted from the example shown in FIG. According to this edge extraction, it is considered that erroneous recognition as shown in FIG. 10 does not occur. Therefore, it is considered that the input data to be rejected can be isolated from the recognition target pattern data by adding the edge connectivity to the feature amount.

Accordingly, the pattern recognition apparatus according to claim 1 is a pattern recognition apparatus for identifying a two-dimensional pattern included in image data, and includes input means for inputting image data and preprocessing for performing preprocessing of the image data. Means, feature extraction means for extracting feature quantities in consideration of edge connectivity from the preprocessed image data, learning pattern storage means for storing feature quantities of previously learned patterns, and the previously learned patterns And an extracted means for comparing the extracted feature quantity with a pattern learned in advance, and an output means for outputting the identification result . Density value gradient calculation means for extracting edge intensity and edge direction component from each pixel of the preprocessed image data, and direction component quantization means for quantizing the direction component , Edge search means for calculating an edge addition value for each pixel, and image data that has been subjected to the preprocessing is divided into a plurality of areas, and a histogram of edge addition values for edge direction components in each area is created. A local direction frequency distribution creating unit that calculates a feature vector from the histogram as the feature amount, and the edge search unit uses one pixel as a pixel of interest and is based on a quantized edge direction component of the pixel of interest. Then, a connected edge candidate pixel is determined from pixels adjacent to the target pixel, and a connected edge pixel having a density gradient value closest to the target pixel is determined from the connected edge candidate pixels. The connection is performed on the pixels, and the connection edge candidate pixel is determined and the connection edge pixel is determined in the same manner as described above using the connection edge pixel as the connection target pixel. Prescribed number of connections continuously performed until reaching that, and calculating for each pixel the sum of the edge intensities of all the connected edge pixels connected to the pixel of interest as the edge added value.

The pattern recognition device according to claim 2 is the pattern recognition device according to claim 1, wherein the local direction frequency distribution creating unit divides the image data into a plurality of regions, and each region is quantized for each direction. Then, the sum of the edge addition values is obtained, and the feature vector having the number of dimensions calculated by adding the number of regions to the quantized number of directions is obtained .

The pattern recognition method according to claim 3 is a pattern recognition method for identifying a two-dimensional pattern included in image data, an input step for inputting image data, and a preprocessing step for performing preprocessing of the image data; , A feature extraction step for extracting feature amounts in consideration of edge connectivity from the preprocessed image data, a learning pattern storage step for storing feature amounts of previously learned patterns, and features of the previously learned patterns An identification step for comparing a quantity with the extracted feature quantity, determining an identification result from patterns learned in advance, and an output step for outputting the identification result. A density value gradient calculating step for extracting edge intensity and edge direction component from each pixel of the processed image data; A direction component quantization step, an edge search step for calculating an edge addition value for each pixel, and the preprocessed image data is divided into a plurality of regions, and an edge addition value for an edge direction component in each region And a local direction frequency distribution creation step of calculating a feature vector from the histogram as the feature amount, wherein the edge search step uses one pixel as a pixel of interest and is a quantized edge direction component Based on the above, the connected edge candidate pixel is determined from the pixels adjacent to the target pixel, and the connected edge pixel having the density gradient value closest to the target pixel is determined from the connected edge candidate pixels. Concatenation is performed on the pixels, and the connection edge candidate pixel is determined as the connection target pixel, and the connection edge candidate pixel is determined and connected in the same manner as described above. Continuously performed until reaching the coupling number that defines the connection to the di-pixel, and calculating means calculates for each pixel the sum of the edge intensities of all the connected edge pixels connected to the pixel of interest as the edge sum value To do.

The pattern recognition method according to claim 4 is the pattern recognition method according to claim 3, wherein the local direction frequency distribution creating step divides the image data into a plurality of regions, and each region is quantized for each direction. Then, the sum of the edge addition values is obtained, and the feature vector having the number of dimensions calculated by adding the number of regions to the quantized number of directions is obtained .

A pattern recognition program according to claim 5 is characterized in that the pattern recognition apparatus or pattern recognition method according to any one of claims 1 to 4 is described as a computer program and can be executed. And

A recording medium according to claim 6 is recorded with a pattern recognition program in which the pattern recognition apparatus or pattern recognition method according to any one of claims 1 to 4 is executable by a computer. And

  By realizing the above means, the evaluation value representing the connectivity of the edge is added to the feature amount, and the input data to be rejected has a feature amount greatly different from the recognition target pattern data. Pattern recognition for natural images can be performed with high accuracy.

  According to the present invention, the difference between the recognition target pattern data and the other input data can be increased by using the feature amount in which the edge connectivity evaluation value is added to the edge strength. It is possible to provide a pattern recognition apparatus, a pattern recognition method, a pattern recognition program, and a recording medium on which a pattern recognition program is recorded for accurately identifying the included two-dimensional pattern.

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

  FIG. 1 is a block diagram showing a configuration of a pattern recognition apparatus according to an embodiment of the present invention, and FIG. 2 is a flowchart of a pattern recognition method according to the embodiment of the present invention.

  1 and 2, first, in S <b> 1, the image data input unit 11 inputs image data such as a natural image captured by a digital camera or the like and transmits the image data to the preprocessing unit 12.

  In S2, the preprocessing unit 12 performs preprocessing. For example, from the input digital image data, some partial areas where characters are supposed to exist are cut out at appropriate sizes and positions, and density value normalization (for example, density value average is 0 and variance is 1 normal) Or processing such as noise removal, and transmits the processed data to the feature extraction unit 13.

  In S <b> 3, the feature extraction unit 13 extracts a feature amount considering the edge connectivity from the input digital image data, and transmits the feature amount to the identification unit 14.

  In S4, the identification unit 14 compares the feature quantities of several patterns stored in the learning pattern storage unit 15 with the transmitted feature quantities, and identifies from the stored patterns based on the comparison results. The result is determined, and the identification result is transmitted to the output unit 16.

  In S5, the output unit 16 outputs the identification result calculated by the identification unit 14 and ends the operation.

  Next, the configuration of the arithmetic processing unit and the process execution method in the feature extraction unit 13 will be described in detail with reference to FIGS. FIG. 3 is a block diagram showing the configuration of the arithmetic processing unit, and FIG. 4 is a flowchart of a process execution method.

  3 and 4, first, in S6, the density value gradient calculation unit 21 calculates the density value gradient (edge) at each pixel point of the input image data. As a result, the magnitude (edge strength) of the density value gradient and the direction component are calculated. The actual calculation can be performed using, for example, a Sobel filter, a Roberts filter, or the like. An example of calculating the magnitude of the concentration value gradient is shown in FIG. In FIG. 5, the darker the density (black), the greater the density value gradient (the higher the edge strength).

  In S <b> 7, the direction component quantization unit 22 quantizes the direction component of the density value gradient in four directions, that is, upper and lower directions, lower left and upper right directions, left and right directions, and upper left and lower right directions.

  In S <b> 8, the edge search unit 23 sequentially searches for the pixel closest to the gradient of the own pixel with respect to each pixel point. The conceptual diagram is shown in FIG. In FIG. 6, the gradient direction of the pixel of interest is quantized in the lower left and upper right directions, and the pixel closest to the gradient size of the own pixel is searched in two directions.

  Here, the search method will be described in detail based on the flowchart of FIG. First, in S10, variables i (image width) and j (image height) representing pixel positions are initialized to 1.

  In S11, the pixel at the position (i, j) is set as the target pixel.

  In S12, the magnitude P of the density value gradient of the target pixel is recorded.

  In S13, the connection number n, which is a variable indicating the number of connections, is initialized with 1.

  In S <b> 14, “connected pixel of interest” is defined in order to sequentially execute the connection process. This “concatenated pixel of interest” is for recording the position of the pixel to be linked from the pixel of interest as a starting point, and is prepared for two pixels for searching in two directions.

  In S15, a connection edge candidate pixel to be connected next is determined among the 8 pixels in the vicinity of the connection target pixel. For example, the determination method is set in a direction other than the density value gradient direction as shown in FIG. 8 based on the density value gradient direction of the target pixel. In FIG. 8, a white circle represents a connection target pixel, and a black circle represents a connection edge candidate pixel. There are two types for the purpose of searching in two directions.

  In S16, among the connected edge candidate pixels, a pixel whose density value gradient is closest to P is determined as a connected edge pixel and connected to this.

  In S17, the connection edge pixel is set as a new connection target pixel.

  In S18, it is determined whether or not the number n of connections has reached a predetermined maximum number N. If it is determined that the number N has not reached N, the number n of connections is set to n + 1 in S23, and the process returns to S15. A connection edge candidate pixel is determined, and the operations of S15 to S18 are repeated. Note that an appropriate value is set in advance for the maximum number N of connections. If it is determined in S18 that the number of connections n has reached the maximum number of connections N, the process proceeds to S19 in the subsequent stage.

  In S19, it is determined whether i has reached the specified value. If it is determined that i has not reached the specified value, i is set to i + 1 in S22, and the process returns to S11, and the operations of S11 to S19 are repeated again. If it is determined in step S19 that i has reached the specified value, the process proceeds to step S20.

  In S20, it is determined whether j has reached a specified value. If it is determined that j has not reached the specified value, j is set to j + 1 in S21, and the process returns to S11, and the operations of S11 to S20 are repeated again. If it is determined in step S20 that j has reached the specified value, the operation is terminated.

  By the above edge search processing, a set of pixels connected to the target pixel can be obtained, and an edge addition value obtained by adding the magnitude values of the density value gradients of all the pixels connected to the target pixel is calculated. Then, edge addition values are calculated for all pixel points.

  3 and 4, in S9, the local direction frequency distribution creation unit 24 creates a local gradient direction component histogram based on the edge addition value. Specifically, as shown in FIG. 9, the image is divided into several partial areas, and the sum of the edge addition values is obtained for each of the four directions in each area block. With the above processing, a 4 × (region block number) -dimensional feature vector is calculated.

  In the above description, the direction component quantization number has been described as four directions. However, it is possible to expand the number of directions component to eight directions or 16 directions by changing the connected edge candidate pixel determination method of the edge search unit 23.

  Next, processing in the identification unit 14 will be described. The identification unit 14 performs processing using a subspace method.

  This subspace method is a kind of statistical method for determining categories to be classified through projection onto a subspace formed from the distribution of feature vector components. In this case, for the eigenvector calculation of the vector component to be converted, for example, KL analysis by Karhunen-Loeve transform which is a quantization algorithm is employed. As typical techniques in the subspace method, the CLAFIC method and the average learning subspace method are known. ALSM belongs to an adaptive learning approach that also takes into account opposing categories, and the learning of category boundaries is advanced by re-adjusting the space repeatedly so as to minimize the misrecognition of a predetermined training pattern.

  First, sample data of a pattern to be identified is input in advance, a partial space representing the pattern is calculated and stored in the learning pattern storage unit 15. The subspace can be obtained by calculating the eigenvalue and eigenvector of the covariance matrix of the feature vector obtained from the identification pattern.

  At the time of identification, the partial pattern data of all patterns is called from the learning pattern storage unit 15, and the pattern is identified based on the distance from these partial spaces. The distance is measured by a simple Euclidean distance in the feature space, and the pattern of the subspace with the closest distance is used as the identification result.

  The basic principle of the subspace method is described in detail in Kenichiro Ishii, Nobuyoshi Ueda, Eisaku Maeda, Hiroshi Murase, “Easy-to-understand pattern recognition”, Ohmsha (1998), etc. In addition to the subspace method, for example, a nearest neighbor method, Fisher's linear discriminant method, support vector machine, neural network, kernel nonlinear subspace method, or the like may be used.

  The present invention also provides a storage medium in which a program code of software for realizing the functions of the above-described embodiments is recorded to a system or apparatus, and a program in which the CPU (MPU) of the system or apparatus is stored in the storage medium. This can also be realized by reading and executing the code. In that case, the program code itself read from the storage medium realizes the functions of the above-described embodiment, and a storage medium storing the program code, for example, a CD-ROM, DVD-ROM, CD-R, CD-RW, MO, HDD, etc. constitute the present invention.

The block diagram which shows the structure of a pattern recognition apparatus. The flowchart of a pattern recognition method. The block diagram which shows the arithmetic processing unit structure of the feature extraction part. The flowchart of the execution method of the process of the characteristic extraction part. The figure which shows the example which computed the magnitude | size of the density | concentration value gradient. The conceptual diagram which searches a pixel sequentially. The flowchart which searches a pixel sequentially. The figure which shows the example of the determination method of the next connection edge candidate pixel. The figure which shows the example of creation of the histogram of a local gradient direction component. The figure which shows the example which misrecognizes with a complicated character when a complicated texture is input. The figure which shows the example which performed the edge extraction process of FIG.

Explanation of symbols

DESCRIPTION OF SYMBOLS 11 ... Image data input part 12 ... Pre-processing part 13 ... Feature extraction part 14 ... Identification part 15 ... Learning pattern memory | storage part 16 ... Output part 21 ... Density value gradient calculation part 22 ... Direction component quantization part 23 ... Edge search part 24 ... Local direction frequency distribution generator

Claims (6)

A pattern recognition apparatus for identifying a two-dimensional pattern included in image data,
Input means for inputting image data;
Preprocessing means for performing preprocessing of the image data;
Feature extraction means for extracting feature quantities considering the connectivity of edges from the preprocessed image data;
Learning pattern storage means for storing feature quantities of previously learned patterns;
An identification unit that compares the feature quantity of the previously learned pattern with the extracted feature quantity, and determines an identification result from the previously learned pattern;
Output means for outputting the identification result ,
The feature extraction means includes
Density value gradient calculating means for extracting edge intensity and edge direction component from each pixel of the preprocessed image data;
Direction component quantization means for quantizing the direction component;
Edge search means for calculating an edge addition value for each pixel;
The pre-processed image data is divided into a plurality of regions, a histogram of edge addition values for edge direction components in each region is created, and a local direction frequency distribution that calculates a feature vector from the histogram as the feature amount Creating means, and
The edge search means sets one pixel as a target pixel, determines a connected edge candidate pixel from pixels adjacent to the target pixel based on a quantized edge direction component of the target pixel,
Among the connected edge candidate pixels, the one having the density gradient value closest to the target pixel is determined as a connected edge pixel,
Concatenation is performed for the determined connected edge pixel, and the connected edge pixel is used as a connected pixel of interest, and the connection edge candidate pixel is determined and connected to the connected edge pixel in the same manner as described above until the number of connections specified is reached. To
The sum of edge strengths of all connected edge pixels connected to the target pixel is calculated for each pixel as the edge addition value.
A pattern recognition apparatus. The local direction frequency distribution creating means divides image data into a plurality of regions,
  In each region, find the sum of the edge addition values for each quantized direction,
  A feature vector having a number of dimensions calculated by adding the number of regions to the quantized number of directions is obtained.
  The pattern recognition apparatus according to claim 1. A pattern recognition method for identifying a two-dimensional pattern included in image data,
  An input step for inputting image data;
  A preprocessing step of performing preprocessing of the image data;
  A feature extraction step of extracting feature quantities considering the connectivity of edges from the preprocessed image data;
  A learning pattern storage step for storing a feature amount of a previously learned pattern;
  An identification step of comparing the feature quantity of the previously learned pattern with the extracted feature quantity, and determining an identification result from the previously learned pattern;
  An output step of outputting the identification result,
  The feature extraction step includes
  A density value gradient calculating step for extracting an edge strength and an edge direction component from each pixel of the preprocessed image data;
  A direction component quantization step for quantizing the direction component;
  An edge search step for calculating an edge addition value for each pixel;
  The pre-processed image data is divided into a plurality of regions, a histogram of edge addition values for edge direction components in each region is created, and a feature vector is calculated as the feature quantity from this histogram. A creation step,
  The edge search step uses one pixel as a target pixel, determines a connected edge candidate pixel from pixels adjacent to the target pixel based on a quantized edge direction component,
  Among the connected edge candidate pixels, the one having the density gradient value closest to the target pixel is determined as the connected edge pixel,
  Concatenation is performed on the determined connected edge pixel, and the connected edge pixel is used as a connected pixel of interest, and the connection edge candidate pixel is determined and connected to the connected edge pixel in the same manner as described above until the number of connections specified is reached. To
  The sum of edge strengths of all connected edge pixels connected to the target pixel is calculated for each pixel as the edge addition value.
  A pattern recognition method characterized by the above. The local direction frequency distribution creating step divides the image data into a plurality of regions,
  In each region, find the sum of the edge addition values for each quantized direction,
  A feature vector having a number of dimensions calculated by adding the number of regions to the quantized number of directions is obtained.
  The pattern recognition method according to claim 3. A pattern recognition program characterized in that the pattern recognition apparatus or pattern recognition method according to any one of claims 1 to 4 is described as a computer program and can be executed. A recording medium on which is recorded a pattern recognition program in which the pattern recognition apparatus or pattern recognition method according to any one of claims 1 to 4 is executable by a computer.