JP3248965B2 - Binary threshold value determination device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a neural network having a learning function for determining a binarization threshold for binarizing a multivalued image in order to obtain an input image of a character recognition device. The present invention relates to a binarization threshold value determining device used.

[0002]

2. Description of the Related Art In general, a character recognition apparatus recognizes a binary image, and in the case of a multivalued image, it is necessary to input the binary image. Here, the recognition performance of the character recognition device depends on the quality of the binary image. Therefore, if a binarization processing unit for obtaining a high-quality binary image is added at the preceding stage of the character recognition device, the recognition performance of the character recognition device can be improved. It can be said that the performance can be improved.

[0003] Therefore, as a method of determining a binarization threshold value for binarizing a multi-valued image, many methods have been conventionally proposed and announced. To give a few examples, first, a paper "Automatic threshold selection based on discrimination and least square criterion" (IEICE Transactions D Vol. J63 No.4 pp.3
49-359). This is because the problem of threshold selection is grasped in a general basic framework, and the 0th-order and 1st-order cumulative moments of the density histogram are considered from the viewpoint of a criterion for maximizing the degree of separation at the density level of the class to be separated. In this case, only an optimal method is used in the sense of least-squares approximation of a gray-scale image.

[0004] Second, a paper “Threshold Determination Method Focusing on Average Number of Adjacent Numbers” (IEICE Transactions D-II Vol.J
73 No. 3 pp. 360-366). This defines the “average number of neighbors” as a measure of the goodness of the connected components when binarizing the grayscale image (multi-valued image) while changing the threshold value,
The value at which this measure becomes maximum is determined as an appropriate threshold value. At this time, in order to prevent the amount of calculation from increasing in proportion to the number of thresholds to be changed, the average number of neighbors is calculated by a combination of rank filter and histogram processing with the amount of calculation independent of the number of thresholds to be changed. Has been devised so that can be calculated at high speed.

Third, the paper "Proposal of a new method of calculating a binary threshold value of a noisy gray image (local histogram difference method)" or "a method of calculating a threshold value of a noisy gray image (local histogram difference method)" ) "(1990 IEICE Autumn National Convention D-365,366). This utilizes the difference between the high density area and the low density area per unit area histogram, that is, the difference between the normalized histograms.
The true threshold value can be calculated without regard to the noise level or the pattern rate in the calculation of the threshold value.

Further, as an example of using a neural network for binarizing a multi-valued image, see the paper "Relaxed Neural Network Model for Optimizing Image Binarization and Its Realization by Parallel Computer" (Transactions of the Institute of Electronics, Information and Communication Engineers). D-II Vol.J74
No.6 pp.678-687) using a neural circuit model. This is to optimize the binarization by minimizing the energy function defined by the grayscale values of the pixels and the error between the binarized pixels in all the local neighborhood systems on the image.

[0007]

However, these methods have the following disadvantages. For example, in the discriminant analysis method and the local histogram method, the accuracy of the determination of the binarization threshold value is not sufficient as preprocessing of the character recognition device, and only a low-quality binary image may be obtained depending on the original. Further, in the method using the average number of neighbors, processing is performed on all of the binarized images obtained using the respective binarization thresholds (adaptive threshold method), so that the processing time is long. Further, in the method using the neural circuit model, since the values of several pixels near the point of interest are input, one-dot noise is likely to occur, which is not very suitable as preprocessing of the character recognition device.

As described above, the preprocessing of the character recognition apparatus is still insufficient by any of the methods. In particular,
In the discriminant analysis method and the local histogram method, only a part of the information obtained from the density histogram is used, and therefore it is considered that the accuracy of determining the binarization threshold value is not sufficient.

[0009]

In the invention SUMMARY OF THE INVENTION 請 Motomeko 1, provided with calculating means for calculating a mean value and a variance value of the density of each pixel of the multivalued image, a neural network, the obtained density of the average value And an input layer that inputs a variance value and an output layer that outputs an optimal binarization threshold value for the input multi-valued image.

[0010] In addition, in the invention of claim 2, wherein, in the neural network, 2 value calculated from the value output from the output layer is given to the input layer of the average value and variance value of the density obtained from the multi-level image Learning control means for learning the neural network with an erroneous recognition rate of the character recognition device at the time of binarization using the threshold value as an error.

According to a third aspect of the present invention, there is provided a means for creating a density histogram from each pixel value of the multi-valued image, and a calculating means for calculating an average value and a variance value of the density of each pixel of the multi-valued image. The network is provided with an input layer for inputting the density histogram, the average value and the variance of the density, and an output layer for outputting an optimal binarization threshold value for the input multi-valued image.

[0012] In addition, in the invention of claim 4, wherein, in the neural network, the concentration histogram created from the multi-level image, is output from the output layer is given to the input layer of the average value and the variance of the determined concentration Learning control means is provided for learning the neural network with the error recognition rate of the character recognition device when binarized using the binarization threshold value calculated from the value as an error.

Further, in the invention according to the fifth aspect, means for creating a two-dimensional histogram relating to the density of each pixel of the multi-valued image and the inclination of each pixel is provided, and the neural network is provided with a means for generating the two-dimensional histogram. An input layer for inputting a frequency value on each coordinate and an output layer for outputting an optimal binarization threshold value for the input multi-valued image are provided.

In addition, in the invention according to claim 6 , the frequency value on each coordinate of the two-dimensional histogram created from the multi-valued image is given to the input layer to the neural network, and the value output from the output layer is given. A learning control means is provided for learning the neural network with an erroneous recognition rate of the character recognition device at the time of binarization using the binarization threshold calculated as an error.

According to the seventh aspect of the present invention, the third or fourth aspect is provided.
In the described invention, the density histogram is normalized by the average value and the variance value of the multi-valued image.

[0016]

[Action] In the invention 請 Motomeko 1, wherein, because it determines the binarization threshold by inputting the average value and the variance value of the density of each pixel of the multivalued image to the neural network, more information And it is possible to determine the binarization threshold with high accuracy. According to the third and fourth aspects of the present invention, the binarization threshold value is determined by inputting the average value and the variance value of the density of each pixel of the multivalued image together with the entirety of the density histogram to the neural network. It is possible to determine the binarization threshold with high accuracy. Also in the inventions according to claims 5 and 6 , since a two-dimensional histogram in which the inclination information of each pixel is added to the density histogram is input to the neural network to determine the binarization threshold value, binarization with higher accuracy is performed. The threshold can be determined. According to the seventh aspect of the present invention, the density histograms handled by these inventions are normalized by the average value and the variance of the multi-valued image, so that variations in the density histogram due to differences in scanners and the like are reduced. .

[0017]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A first embodiment of the present invention will be described with reference to FIGS. The processing device of the present embodiment is for converting multi-valued image information read by the image reading device (scanner) 1 into binary information and providing the binary information to the character recognition device 2. Is configured as a binarization threshold value determination device 4 that outputs a threshold value for binarization to a binarization processing unit 3 that performs the binarization. First, the image reading apparatus 1 reads a document image as multivalued image information in which each pixel has 256 gradations, for example. The read multilevel image information is stored in the multilevel image memory 5. The binarization processing unit 3 binarizes the multivalued image in the multivalued image memory 5 using the binarization threshold value determined by the binarization threshold value determination device 4.
More specifically, a pixel having a pixel value larger than the binarization threshold is set to “1”, and the other pixels are set to “0” as binary image information, which is output to the character recognition device 2. The character recognition device 2 performs line cutout, character cutout, matching, language processing, and the like on the input binary image information, and performs processing for converting a character image on the binary image into a character code. The processing result by the character recognition device 2 is output to a result output unit 6.
Is output through Specifically, the recognition result is displayed on the display 7 or the recognition result is stored in the file 8 so that the user can perform the correction processing of the recognition result.

Thus, the binarization threshold value determining device 4 of the present embodiment comprises a density histogram creating section (creating means) 9 and a neural network 10 having a learning function. The density histogram creating section 9 creates a density histogram from each pixel value of the multi-valued image. The neural network 10 calculates a binarization threshold based on the created density histogram.

The details of the neural network 10 will be described with reference to FIG. The neural network 10 has a three-layer structure including an input layer 11, an intermediate layer 12, and an output layer 13, for example.
The input layer 11 composed of 56 neuron units (shown by ○) is set so that the frequency value at each density value of the density histogram is input. Each neuron unit of the input layer 11 and five neuron units of the intermediate layer 12 are coupled with coupling coefficients. The output layer 13 is composed of one neuron unit and outputs an optimum binarization threshold Th as a predetermined calculation result. Here, each neuron unit of the intermediate layer 12 and the neuron unit of the output layer 13 are connected by a connection function of tanh (X) as shown on the right side in FIG. Is set to be output. Here, assuming that the output of the output layer 13 is out, the binarization threshold Th is obtained as Th = 128 (out + 1) (1).

In such a configuration, first, it is necessary to perform a learning process of the neural network 10. As the learning data, multivalued image information and an erroneous recognition rate obtained from the character recognition device 2 when the multivalued image information is binarized by each binarization threshold are used. In general, in the character recognition device 2, the recognition performance is deteriorated even when any of a character image is blurred or crushed. Therefore, the binarization threshold and the erroneous recognition rate obtained by the character recognition device 2 are as shown in FIG. The relationship shown in FIG.

At the time of learning, a density histogram is input to the input layer 11, and a function as shown in FIG. 3 obtained from the corresponding multi-valued image is used as an error function in the output layer 13 using the equation (1). An error is calculated from the output out, and learning of the neural network 10 having the configuration shown in FIG. 2 is performed by the back propagation method. Such control is performed by learning control means (not shown).

When the neural network 10 after learning is used, a density histogram obtained from a multi-valued image to be binarized is input to the input layer 11, and the value output from the output layer 13 is used to calculate the expression (1). The binarization threshold value is determined in accordance with the following formula, and the binarization processing unit 3 binarizes the binary threshold value using the determined binarization threshold value, and causes the character recognition device 2 to input the binarization threshold value.

As described above, according to the present embodiment, since the entire density histogram is input to the neural network 10 to determine the binarization threshold, much information is obtained from the density histogram. This means that the binarization threshold value is determined with high accuracy, and the processing speed for that can be realized at a sufficient speed.

Next, a second embodiment of the present invention will be described with reference to FIG. The same parts as the parts shown in the above-mentioned embodiment are denoted by the same reference numerals (the same applies to the following embodiments).
This embodiment is designed to further improve the accuracy of the determination of the binarization threshold value. In addition to the embodiment, the average value and the variance value of the density of each pixel of the multi-valued image are taken into account. Is configured. Such mean and variance values are
Referring to FIG. 1, it is calculated in the density histogram creating section 9.

The neural network 10 of the present embodiment
Is input with the average value and the variance value of the density of each of these pixels together with the density histogram. The input layer 11 has two neuron units added, and a total of 2
It is composed of 58 neuron units. More specifically, the average value and the variance of the density histogram obtained from the multi-valued image are obtained, and the distribution of the density histogram obtained from the multi-valued image is assumed to be a normal distribution. Is converted into a standard normal distribution of 1 so that a normalized density histogram is input. Such normalization of the density histogram makes it possible to eliminate variations in the density histogram due to differences between the image reading apparatuses 1 and the like.

Here, since the average value and the variance of the density of each pixel of the multi-valued image greatly affect the accuracy of the determination of the binarization threshold value, the neural network 10 of this embodiment is different from the density histogram. Is formed to form a hierarchical network. That is, the addition unit 14 that adds only the bias y having a variable bias value to the output layer 13.
And a multiplication unit 15 for multiplying the output of the addition unit 14 by the variance value is provided. The average value and the output of the multiplication unit 15 weighted by the variable coupling coefficient X are added to form a binarization threshold. An addition unit 16 that outputs the value Th is provided. Between the output layer 13 and the addition unit 14, between the addition unit 14 and the multiplication unit 15, between the variance value unit and the multiplication unit 15 in the input layer 11, and between the average value unit in the input layer 11 and the addition. The connection weight with the unit 16 is always set to one.

In such a configuration, when learning,
The density histogram, the average value of the density, and the variance value are input to the input layer 11, and the function shown in FIG. 3 obtained from the corresponding multi-valued image is output from the output layer 13 as an error function.
An error is calculated from the threshold value, and learning of the neural network 10 having the configuration as shown in FIG. 4 is performed by the back propagation method.

When the trained neural network 10 is used, a density histogram, a density average value and a variance value obtained from a multivalued image to be binarized are input to the input layer 11 and output from the output layer 13. The binarization threshold value is determined from the threshold value, and the binarization processing unit 3 binarizes the binarization value using the determined binarization threshold value, and causes the character recognition device 2 to input the binarization threshold value.

Further, a third embodiment of the present invention will be described with reference to FIG. In the present embodiment, the two-dimensional histogram as shown in FIG. 5 relating to the density of each pixel and the inclination of each pixel of the multi-valued image is created by the density histogram creating unit 9 shown in FIG. Things. The configuration of the neural network 10 used in the present embodiment is basically similar to that shown in FIG. 2, but input of such a two-dimensional histogram (frequency value at each coordinate).
The input layer 11 receiving the input data is composed of 256 × 256 = 65536 neuron units.

That is, in this embodiment, in order to further improve the accuracy of the determination of the binarization threshold, in addition to the density histogram, the inclination information of each pixel is used as an input. The two-dimensional histogram as shown in FIG. 5 is created by calculating the value and the slope and incrementing the frequency value of the corresponding coordinates of the two-dimensional histogram. For example, the coordinate a in FIG. 5 indicates that the number of pixels having the inclination 130 and the density 130 is 20 (frequency value = 20). When the pixel coordinates are (x, y) (pixel value f (x, y)), the inclination g of each pixel is the value of four pixels (f (x, y−1), f (x, y,
y + 1), f (x-1, y), f (x + 1, y)), g = [4f (x, y)-{f (x-1, y) + f (x + 1, y) + f ( x, y-1) + f (x, y + 1)}] / 8 + 128 (2)

In such a configuration, when learning,
A two-dimensional histogram is input to the input layer 11, and an error is calculated from the binarization threshold value output from the output layer 13 using the function shown in FIG. 3 obtained from the corresponding multi-valued image as an error function. Learning of the neural network 10 is performed by the propagation method.

When the trained neural network 10 is used, a two-dimensional histogram is input to the input layer 11, a binarization threshold is determined from a value output from the output layer 13, and the determined binarization threshold is determined. The values are binarized by the binarization processing unit 3 and input to the character recognition device 2.

[0033]

According to the invention 請 Motomeko 1, wherein according to the present invention was to determine the allowed input binary threshold value the average value and variance value of the density of each pixel of the multivalued image in a neural network Therefore, more information can be obtained, and the binarization threshold value can be determined with high accuracy. Further, according to the third and fourth aspects of the present invention, the binarization threshold value is determined by inputting the average value and the variance value of the density of each pixel of the multi-valued image together with the entire density histogram to the neural network. So that more accurate 2
Binarization threshold decision is possible, likewise, claim 5,
According to the invention described in the sixth aspect, the two-dimensional histogram in which the inclination information of each pixel is added to the density histogram is input to the neural network to determine the binarization threshold value. The threshold can be determined. According to the seventh aspect of the invention, the density histograms handled in these inventions are normalized by the average value and the variance value of the multi-valued image. Become.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a first embodiment of the present invention.

FIG. 2 is a schematic diagram showing the configuration of the neural network.

FIG. 3 is a characteristic diagram of a binarization threshold value-misrecognition rate.

FIG. 4 is a schematic diagram of a neural network configuration showing a second embodiment of the present invention.

FIG. 5 is an explanatory diagram of a two-dimensional histogram showing a third embodiment of the present invention.

[Explanation of symbols]

 2 Character recognition device 9 Creation means or calculation means 10 Neural network 11 Input layer 13 Output layer

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-4-297961 (JP, A) JP-A-5-328133 (JP, A) Babaguchi et al., "Experimental image binarization using a connectionist model" Examination ”, IEICE Transactions D-II, Vol. J73-D-II, No. 8, pp. 1281-1287, August 1990 (58) Fields investigated (Int. Cl. ⁷ , DB name) G06K 9/38 G06N 3/00 G06T 1/00 G06T 5/00

Claims (7)

(57) [Claims] 1. A binarization threshold value determining device for determining a threshold value for converting a multi-valued image into a binary image as an input image to a character recognition device using a neural network having a learning function. Each pixel value of the multi-valued image
Means for creating a density histogram, the <br/> and calculating means for calculating a mean value and a variance value of the density of each pixel of the multivalued image provided from said neural network, conc obtained
An input layer for inputting a degree histogram and an average value and a variance value of density, and an output layer for outputting an optimal binarization threshold value for an input multi-valued image; Decision device. 2. A binarization threshold value determining device for determining a threshold value for converting a multi-valued image into a binary image as an input image to a character recognition device using a neural network having a learning function. Each pixel value of the multi-valued image
Means for creating a density histogram, the <br/> and calculating means for calculating a mean value and a variance value of the density of each pixel of the multivalued image provided from said neural network, conc obtained
An input layer for inputting the average value and variance value of the density histogram and the density, an output layer for outputting an optimal binarization threshold value for the input multi-valued image, and an average value of the density obtained from the multi-valued image And a variance value given to the input layer and binarized using a binarization threshold calculated from a value output from the output layer. A binary threshold value determining device, comprising: learning control means for performing learning. 3. A binarization threshold value determining device for determining a threshold value for converting a multivalued image into a binary image as an input image to a character recognition device using a neural network having a learning function, A means for creating a density histogram from each pixel value of the multi-valued image, and a calculating means for calculating an average value and a variance value of the density of each pixel of the multi-valued image are provided. An apparatus for determining a binarization threshold value, comprising: an input layer for inputting an average value and a variance value of density; and an output layer for outputting an optimal binarization threshold value for an input multi-valued image. 4. A binarization threshold value determining device for determining a threshold value for converting a multivalued image into a binary image as an input image to a character recognition device using a neural network having a learning function, A means for creating a density histogram from each pixel value of the multi-valued image, and a calculating means for calculating an average value and a variance value of the density of each pixel of the multi-valued image are provided. An input layer for inputting the average value and the variance value of the density, an output layer for outputting an optimal binarization threshold value for the input multi-valued image, a density histogram created from the multi-valued image, a calculated density Erroneous recognition rate of the character recognition apparatus when the average value and the variance value are given to the input layer and binarized using a binarization threshold value calculated from a value output from the output layer And a learning control means for performing the learning of the neural network with the error as an error. 5. A binarization threshold value determining device for determining a threshold value for converting a multivalued image into a binary image as an input image to a character recognition device using a neural network having a learning function, A means for creating a two-dimensional histogram relating to the density of each pixel and the inclination of each pixel of the multi-valued image; and inputting a frequency value on each coordinate of the created two-dimensional histogram to the neural network. An apparatus for determining a binarization threshold value, comprising: an input layer; and an output layer for outputting an optimal binarization threshold value for an input multi-valued image. 6. A binarization threshold value determining device for determining a threshold value for converting a multivalued image into a binary image as an input image to a character recognition device by using a neural network having a learning function, A means for creating a two-dimensional histogram relating to the density of each pixel and the inclination of each pixel of the multi-valued image; and inputting a frequency value on each coordinate of the created two-dimensional histogram to the neural network. An input layer, an output layer that outputs an optimal binarization threshold value for the input multi-valued image, and a frequency value on each coordinate of a two-dimensional histogram created from the multi-valued image is given to the input layer. The learning of the neural network is performed using the error recognition rate of the character recognition device as an error when binarizing using the binarization threshold calculated from the value output from the output layer. And a learning control means for performing the threshold value determination. 7. The binarization threshold determining device according to claim 3 or 4, characterized in that so as to normalize the mean and the variance of the density histogram multivalued image.