JPH06113139A - Picture binarization device - Google Patents

Abstract

PURPOSE:To provide a picture binarization device where an appropriate threshold based on the histogram of lightness is decided and an input picture can satisfactorily and speedily be binarized through the use of the threshold even if illumination unevenness occur in the input picture. CONSTITUTION:A character picture is image-picked up by an image sensor 1, is digitalized in an A/D converter 3 and is stored in a frame memory 5. The inputted character picture is separated into plural partial pictures by a window circuit 7, a block separation circuit 9 and a block selection circuit 11. Then, a notice partial picture block and plural adjacent partial picture blocks adjacent to the notice partial picture block are selected from the plural partial pictures. A histogram unit 13 generates histograms based on lightness as to the notice partial picture block and the adjacent partial picture blocks. The thresholds of the respective partial picture blocks are outputted from data of the histograms by using a neural network 18 and the input picture is binarized by the threshold.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a binarizing device suitable for digitally converting an image signal.

[0002]

2. Description of the Related Art Generally, in a copying machine or a facsimile machine, in order to make a character image drawn on paper clear, a photographed image having gradation is binarized. Normally, a threshold value is determined by a histogram representing the brightness distribution of the input image, and if each pixel of the input image is higher than the threshold value, it is determined as “white”, and if it is lower, “black”, and an output pixel is generated. .

However, the correspondence between the histogram and the threshold value suitable for binarization is not clear. Further, the shape of the histogram to be used as a reference differs greatly due to the difference in illumination and the difference in input character image. As a solution to this problem, Dr. Otsu has published a paper "Automatic threshold selection method based on discrimination and least squares criterion" (IEICE Transactions D, Vol.J63-D No.4 p.
p.349-356), we propose a method to determine a statistically optimum threshold value based on the histogram of the brightness of the input image.

On the other hand, Babaguchi et al., "Experimental study of image binarization by connection model" (IEICE Transactions D-2 Vol.J73-D-2 No.8 pp.1281-1287) )
In, we take a histogram of the brightness of the entire input image, input it to a hierarchical neural network that has the same number of input units and output units as the number of dimensions of the histogram, and use the output unit with the strongest firing as a threshold to input the binary image. I am proposing a method to realize it.

[0005]

In both of the above-mentioned two methods, an input image area having a certain size is required when creating a brightness histogram of the input image. However, when a large image area is input to create a histogram in this way, an accurate histogram cannot be obtained and an appropriate threshold value cannot be determined if there is illumination unevenness in the input image. .

Therefore, according to the present invention, an accurate histogram can be created even when illumination unevenness occurs in the input image, and an appropriate threshold value is determined based on the histogram.
An object of the present invention is to provide an image binarization device that can quickly binarize an input image using this threshold value.

[0007]

In order to achieve the above object, the present invention provides an image pickup means for photoelectrically converting an image to take out as an electric signal, and an image signal from the image pickup means, which is divided into a plurality of parts. Image signal dividing means for generating an image; block selecting means for selecting a target partial image block and a plurality of adjacent partial image blocks adjacent to the target partial image block from the plurality of partial images; and the target partial image. Histogram generating means for generating a histogram based on lightness for each block and adjacent partial image blocks, a neural network for outputting the threshold value of the partial image block of interest from the data of the histogram generated by the histogram generating means, and output from the neural network Based on a threshold that is Of is obtained by and a binarizing means for binarizing at least a portion of the pixel.

[0008]

In the image binarization apparatus having the above-described structure, the input image is separated into a plurality of partial images, and the partial image of interest and a plurality of adjacent partial image blocks are extracted from the plurality of partial images. Select adjacent image blocks and
A histogram is generated for each of the attention partial image block and the adjacent partial image block based on lightness, and the partial image block is binarized using a neural network that outputs the threshold value of the attention partial image block from the data of the histogram. It

[0009]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Prior to the description of the embodiments of the present invention, a method of setting a threshold value for a histogram of brightness of an image will be described. First, divide the input image into multiple partial images,
The characteristic part of the divided partial image is extracted. The lightness is used to characterize this partial image. For example, when binarizing a character image input as a grayscale image, in order to eliminate the influence of illumination unevenness, not the whole image but the whole image is divided into partial images, and a histogram of the brightness of each partial image is displayed. create.

FIG. 9 shows an example of a brightness histogram of a partial image of a character image having a size of 32 × 32. FIG. 9 (a)
Is an example of a histogram of the brightness of an area including a character of a printed matter printed in black on a white background. FIG. 9B is a histogram of the lightness of a portion of a printed matter equivalent to that of FIG. As is clear from these figures, the white background region has a narrow lightness distribution range and a sharp peak.

On the other hand, in the area including the character shown in FIG. 9A, a sharp peak of a white background and an area corresponding to the black color of the character gently spread in a range having a lightness smaller than this peak. By comparing these histograms, the difference between the maximum brightness and the minimum brightness is small, and the ratio of the peak height to the area of the entire histogram is large in the area having only the white background. However, in a region including characters, the difference between the maximum brightness and the minimum brightness is widely distributed, and the ratio of the peak height to the area of the entire histogram is small. If the partial image extracted from the input character image is an area with only a white background, the threshold value is set to a lightness lower than the minimum lightness.

On the other hand, in the case of a region including characters, a threshold value is set at a lightness lower than the maximum frequency and higher than the minimum brightness. In order to set the optimum threshold value for each partial image, it is optimum for the operator to perform the operation while visually observing the binarized image with various set values. In the embodiment of the present invention, a partial image of interest and eight adjacent partial images adjacent to the partial image of interest are extracted from a plurality of partial images, and the histogram data of the brightness of each of the nine partial images is input to the neural network. The partial image is binarized using the output value of the neural network as a threshold.

In another embodiment, as a representative of the characteristics of the partial image, the maximum brightness, the minimum brightness, the maximum brightness of the partial image, and the ratio of the maximum frequency to the number of pixels in the entire histogram are represented. The partial image is input to a previously learned neural network, and the partial image is binarized using the output value of the neural network as a threshold. Furthermore, in another embodiment, a histogram of a plurality of partial images extracted from one or a plurality of sample character images is created,
The maximum brightness and the minimum brightness of each partial image, the brightness of the maximum frequency, and the ratio of the maximum frequency to the number of pixels of the entire histogram are used as learning data, and the optimal threshold value for this partial image is used as teacher data to perform neural network learning.

Embodiments of the present invention will be described in detail below with reference to the drawings. FIG. 1 shows a first image binarization apparatus of the present invention.
It is a figure which shows the structure of an Example. As shown in this figure, first, a character image is captured by the image sensor 1 such as a scanner or a television camera. The output of the image sensor 1 is digitized by the A / D converter 3 and temporarily stored in the frame memory 5. The character image is divided into a plurality of partial image blocks of a predetermined size by the block dividing circuit 9 based on the address information generated by the window circuit 7. Then, the block selection circuit 11
Thus, as shown in FIG. 2, one target partial image block is selected from the plurality of partial image blocks based on a predetermined condition, and eight blocks located around the target partial image block are selected. , Is selected as an adjacent partial image block. The attention partial image block,
The eight adjacent partial image blocks are sequentially read into the histogram unit 13 one by one. This histogram device 1
For each pixel of the read partial image, 3 adds "1" to the address corresponding to the brightness of the pixel in the histogram memory 15. Before the image sensor 10 captures a character image, the histogram memory 15
It is assumed that the contents of all addresses of are initialized to "0".

When the reading of the partial images of one target partial image block and eight adjacent partial image blocks in total of nine blocks from the frame memory 5 is completed, the histogram memory 15 stores the partial images as shown in FIG. 9, for example. Such nine brightness histograms corresponding to the respective partial images are stored. The brightness histogram data stored in the histogram memory 15 is sent to the neural network 18 and the learning data memory 20. Here, the neural network 18 constitutes the hierarchical structure shown in FIG.

In the present embodiment, the input layer unit outputs the input signal to the intermediate layer neuron units 31a-3.
In FIG. 3, the branch terminals 30a to 30c are used to represent the branch terminals 30a to 30c, because the branch terminals 30a to 30c have a function of only being distributed in parallel to 1c. Each of the neuron units 31a to 31c, 32 executes the processing shown in Expression (1).

[0017]

[Equation 1]

Here, xi is the i-th input, Wi is the weight for the i-th input, N is the number of inputs, θ is the threshold value of the neuron unit, and y is the output of the neuron unit. In the neuron units 31a to 31c in the intermediate layer, f is a sigmoid function shown in Expression (2).

[0019]

[Equation 2]

In the present embodiment, the output neuron unit 32 is a linear unit, so that f (x) = x ... (3). In FIG. 4, the neuron units 31a to 31 are shown.
The block diagram of c and 32 is shown. In the function table 46 of the neuron units 31a to 31c in the intermediate layer, the formula (2)
The relationship between x and f (x) is stored. The weight Wi of the equation (1) of each neuron unit 31a to 31c, 32
(I = 1, ..., N) and the threshold value θ are determined by a generalized delta rule learning method described later. However, here, it is assumed that the weights and thresholds of the neuron units 31a to 31c, 32 are predetermined.

Each neuron unit 31a-3 of the intermediate layer
1c via branch terminals 30a to 30c,
From the histogram memory 15, histogram data of brightness of partial images of one target partial image block and eight adjacent partial image blocks in total of nine blocks is input. The input histogram data is sent to the calculator 41 via the input buffer 40. In the calculator 41, the product sum of the frequency data for each brightness and the weight data for the frequency stored in the weight memory 45 is calculated based on the histogram data of the brightness of the partial image for the nine blocks, and The calculation result is stored in the register 43.

Through the above process, the operation unit 41 calculates and stores in the register 43, based on the histogram data of the partial images of one target partial image block and eight adjacent partial image blocks in total. The value stored in the threshold memory 44 is subtracted from the value. As a result, the register 43 has the following formula (1):

[0023]

[Equation 3]

The result obtained by calculating the part of is stored. The value of the sigmoid function corresponding to the value stored in the register 43 is obtained from the function table 46 and output as an output of each of the intermediate layer neuron units 31a to 31c via the output buffer 42. The output neuron unit 32 sequentially reads the outputs of the intermediate neuron units one by one.

First, the output of the intermediate neuron unit 31a shown in FIG.
Enter 2. The output from the output neuron unit 32 is input to the arithmetic unit 41 via the input buffer 40 shown in FIG. The arithmetic unit 41 calculates the product of the output value of the intermediate neuron unit 31a and the first value of the weight memory 45, and the calculation result is stored in the register 43.

Similarly, the output of the intermediate neuron unit 31b is read. The intermediate neuron unit 31b
The output of is input to the arithmetic unit 41 via the input buffer 40. The arithmetic unit 41 calculates the product of the output value of the intermediate neuron unit 31b and the second value of the weight memory 45, and the result is added to the value of the register 43.

The above process is repeated for all neuron units in the hidden layer. After the processing on the outputs of all the hidden layer neuron units is completed, the value stored in the threshold memory is subtracted from the value stored in the register 43. As a result, the register 43 stores the result calculated by the above equation (1). The value of the calculation result is output via the output buffer 42 as the output value of the neural network 18. The output of the neural network 18 is input to the binarizer 19.

Of the character image stored in the frame memory 5, all or some of the pixels of the target partial image block are sequentially read out and input to the binarizer 19. When the brightness of each pixel is larger than the output value of the neural network 18, it is determined to be a white pixel, and a white pixel having a predetermined brightness is stored in a pixel position of the frame memory 24 corresponding to the pixel position of the frame memory 12. To do.

If the brightness of the pixel read from the frame memory 5 is smaller than the output value of the neural network 18, it is determined to be a black pixel and the frame memory 24 corresponding to the pixel position of the frame memory 12 is determined.
A black pixel having a predetermined brightness is stored at the pixel position of. When binarization of a predetermined number of pixels of one target partial image block is completed in this way, all storage areas of the histogram 15 are initialized to 0, and the next target partial image block and eight adjacent image blocks are Is input to the neural network 18 and the threshold value is obtained. The threshold value is used to binarize the partial image of interest.

When binarizing some pixels of the target partial image block, the original target partial image block and the new target partial image block partially overlap each other.
The above process is repeated for all the blocks of the image in the frame memory 5, and all the images are binarized and stored in the frame memory 24. The weights and thresholds of the neuron units 31a to 31c, 32 of the neural network 18 are generalized delta rules devised by David Lamelhardt et al. (“PDP Model / Cognitive Science and Search for Neuron Networks”, Chapter 8, DE Lamel Heart, JL
McClellan, PDP Research Group, translated by Shunichi Amari, Sangyo Tosho, 1989).

First, the image sensor 10 captures one or a plurality of sample images and extracts a plurality of partial sample images for each sample image. For each partial sample image, as described above, the histogram device 13 and the histogram memory 15 determine the histogram data of the brightness of one target partial image and eight adjacent partial images, and the learning data memory 2
Stored in 0. At the same time, the optimum threshold value for this partial sample image is stored in the teacher data memory 22.

That is, 1 is stored in the learning data memory 20.
The brightness histogram data of one target partial image and eight adjacent partial images is stored. In the teacher data memory 22, a binarization threshold value corresponding to the learning data stored in the learning data memory 20 is stored as teacher data. In the generalized delta rule, the learning data stored in the learning data memory 20 and the teacher data stored in the teacher data memory 22

[0033]

[Equation 4]

Each neuron unit 3 so that
The weights and thresholds of 1a to 31c and 32 are changed. Here, yp is the output of the neural network obtained when the learning data vector xp is input. The learning vector is a vector whose element is the frequency for each lightness of the histogram. dp is teacher data corresponding to the learning vector xp. m is the number of learning data.

The update amount ΔW of each weight is obtained by the steepest descent method as shown in equation (6).

[0036]

[Equation 5]

The update amount ΔWi of the weight for the i-th input of the neuron unit in the output layer is

[0038]

[Equation 6]

In addition, j of the neuron unit i in the intermediate layer is
The weight update amount ΔWij for the th input is

[0040]

[Equation 7]

It is Where yip is the output of the intermediate neuron unit i for the learning data vector xp, W
i is the weight of the i-th input of the output neuron unit, x
hp is the h-th element of the learning vector xp.
Since the threshold value of each neuron unit can be regarded as the weight of the terminal to which "-1" is always input to the neuron unit, the update amount of the threshold value of the neuron unit in the output layer is calculated by the equation (7), The threshold value of the neuron unit of is calculated by equation (8).

In learning, first, the weighting parameters of all neuron units are initialized with random numbers, and the learning device 2
3, the equations (7) and (8) are repeatedly executed.
Then, when E in the equation (5) becomes sufficiently small, the processing is terminated, and the obtained weight is stored in the weight memory 45 and the threshold memory 44 in each of the neuron units 31a to 31c and 32 of the neural network 18.

As described above, the image 2 of this embodiment
According to the binarization device, the input image is divided into one partial image block of interest and eight adjacent partial image blocks into nine partial image blocks, and the histogram of the brightness of each partial image is input to the neural network. As a result, the threshold value for binarization is determined, and thus binarization processing can be performed even on an input image with severe illumination unevenness.

Next, a second embodiment of the present invention will be described with reference to FIG. FIG. 5 is a diagram showing the configuration of the second embodiment of the image binarization apparatus of the present invention. The same members as those in the first embodiment are designated by the same reference numerals, and the details are omitted. The second embodiment differs from the first embodiment in that the maximum / minimum detector 1
6 and that the peak detector 17 is provided.

The maximum / minimum detector 16 obtains the maximum brightness and the minimum brightness from the brightness histogram stored in the histogram memory 15. Further, the peak detector 17 has the maximum frequency of lightness from the lightness histogram stored in the histogram memory 15,
The ratio of the maximum frequency count to the pixel count of the partial image is calculated.

By these maximum / minimum detector 16 and peak detector 17, the input image is composed of one target partial image block and eight adjacent partial image blocks, and a histogram of the brightness of nine partial image blocks in total. Then, the maximum, the minimum, the most frequent lightness, and the maximum frequency are obtained. Using these values, the neural network 18 obtains a threshold value, and the threshold value is used to binarize the image in the same manner as in the first embodiment.

According to the second embodiment, a good binarized image can be obtained even with an image having illumination unevenness as in the first embodiment, and the input dimension of the neural network is larger than that in the first embodiment. Can be significantly reduced, so that the scale of the neural network can be reduced and the learning speed can be improved. 6A and 6B are explanatory diagrams for comparing the binarization processing according to this embodiment and the binarization processing according to the conventional discriminant analysis method, and FIG.
6B is an input image having uneven illumination, FIG. 6B is a binarized image according to this embodiment, and FIG. 6C is a binarized image according to a conventional discriminant analysis method.

As is clear from FIG. 6, from the histogram of each block as in this embodiment, the maximum, minimum,
Even when the binarization process is executed using only four feature quantities, that is, the most frequent lightness and the maximum frequency, a good binary value is obtained even in an image with severe illumination unevenness, which was difficult with the conventional discriminant analysis method. A processed image can be obtained. Next, a third embodiment of the present invention will be described with reference to FIG.

FIG. 7 is a diagram showing the configuration of the third embodiment of the image binarizing apparatus of the present invention. The same members as those in the second embodiment are designated by the same reference numerals, and the details are omitted. The third embodiment differs from the second embodiment in that the histogram memory 1
The output from 5 is divided into two, one target partial image block and eight adjacent partial image blocks, and the histogram data of one target partial image block is directly input to the neural network 18. The data of the histogram of the adjacent partial image block is the second data.
As in the embodiment, the maximum / minimum detector 16 and the peak detector 17 are input to the neural network 18 to obtain the maximum, minimum, most frequent lightness, and maximum frequency.

Also in the third embodiment, as in the first and second embodiments, a good binarized image can be obtained even for an image having uneven illumination. Next, a fourth embodiment of the present invention will be described based on FIG. FIG. 8 is a diagram showing the configuration of the fourth embodiment of the image binarizing apparatus of the present invention. The same members as those in the first embodiment are designated by the same reference numerals, and the details are omitted.

The fourth embodiment differs from the first embodiment in that the selector 25 and the two arithmetic units 27a and 27b are provided.
That is the point. This selector 25 calculates a threshold value based on the output of the neural network 2
One of the two computing units 27a and 27b is selected as an appropriate computing unit.

The computing unit 27a says that the partial image contains pixels of only the background, and the neural network 18
When the judgment is made, a value smaller than the minimum brightness of the partial image or a value larger than the maximum brightness is output as the threshold value. On the other hand, when the neural network 18 determines that the partial image includes pixels other than the background, the calculator 27b outputs the value calculated by the statistical method as the threshold value.

In the present embodiment, since the threshold value for binarization is determined by a statistical method when the target partial image block includes a character portion, the influence of the ambiguity of the output value of the neural network can be reduced. it can. The image is binarized in the same manner as in the first embodiment based on the threshold value obtained by the arithmetic unit thus selected.

Also in the fourth embodiment, as in the first embodiment, a good binarized image can be obtained even with an image having uneven illumination. Next, a fifth embodiment of the present invention will be described with reference to FIG. FIG. 9 is a diagram showing the configuration of the fifth embodiment of the image binarizing apparatus of the present invention. The same members as those in the second embodiment are designated by the same reference numerals, and the details are omitted.

The fifth embodiment differs from the second embodiment in that the selector 25 and the two arithmetic units 27a and 27b are provided as in the fourth embodiment. Also in the fifth embodiment, as in the fourth embodiment, a good binarized image can be obtained even with an image having uneven illumination. The present invention is not limited to the above-described embodiments, and it goes without saying that various modifications and applications can be made without departing from the scope of the invention.

[0056]

As described above in detail, according to the present invention, the histogram data of the brightness of the target partial image block in the input image and a plurality of adjacent partial image blocks adjacent to the target partial image block are used. Even if illumination unevenness occurs in the input image, the learned neural network determines an appropriate threshold value based on the histogram. Using this threshold value, the input image is binarized quickly and satisfactorily. It is possible to provide an image binarization device capable of performing the above.

[Brief description of drawings]

FIG. 1 is a diagram showing a configuration of an image binarization apparatus as a first embodiment of the present invention.

FIG. 2 is a diagram for explaining a relationship between a partial image of interest for all images and an adjacent partial image.

FIG. 3 is a diagram showing a configuration of a neural network.

FIG. 4 is a block diagram showing a configuration of each neuron unit.

FIG. 5 is a diagram showing a configuration of an image binarization apparatus as a second embodiment of the present invention.

FIG. 6 is a schematic diagram illustrating an embodiment of the present invention and a conventional method 2;
It is explanatory drawing for comparing a value-ized process.

FIG. 7 is a diagram showing a configuration of an image binarization apparatus as a third embodiment of the present invention.

FIG. 8 is a diagram showing a configuration of an image binarization apparatus as a fourth embodiment of the present invention.

FIG. 9 is a diagram showing a configuration of an image binarization apparatus as a fifth embodiment of the present invention.

10A is an example of a histogram of the brightness of an area including characters of a certain printed matter printed in black on a white background, and FIG. 10B is a printed matter equivalent to that of FIG. It is a histogram of the brightness of the part of the white background only area.

[Explanation of symbols]

1 ... Image sensor, 3 ... A / D converter, 5, 24 ... Frame memory, 7 ... Window circuit, 9 ... Block division circuit, 11 ... Block selection circuit, 13 ... Histogram device, 15 ... Histogram memory, 16 ... Maximum・ Minimum detector, 17 ... Peak detector, 18 ... Neural network, 19 ... Binarizer, 20 ... Learning data memory, 21 ...
Threshold setting device, 22 ... Teacher data memory, 23 ... Learning device,
25 ... Selector, 27a, 27b ... Arithmetic unit, 30a-30
c ... Branch terminal, 31a to 31c ... (Intermediate) neuron unit, 32 ... Neuron unit, 40 ... Input buffer, 41 ... Arithmetic unit, 42 ... Output buffer, 43 ... Register, 44 ... Threshold memory, 45 ... Weight memory, 46 ... function table.

Claims (1)

[Claims] 1. An image pickup means for photoelectrically converting an image to take out as an electric signal, an image signal division means for dividing a predetermined number of image signals from the image pickup means to generate a plurality of partial images, and the plurality of partial images. More specifically, a block selecting means for selecting a target partial image block and a plurality of adjacent partial image blocks adjacent to the target partial image block, and a histogram based on brightness for each of the target partial image block and the adjacent partial image block Histogram generating means, a neural network that outputs the threshold value of the target partial image block from the data of the histogram generated by the histogram generating means, and at least one of the target partial image blocks based on the threshold value output from the neural network. Binarizing means for binarizing a pixel of a part Image binarization apparatus characterized by comprising a.