JP2000244879A - Image information conversion device and method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To decrease the number of classes with no deterioration of accuracy for an image information conversion device/method. SOLUTION: A class conversion rule generation part 10 checks the frequency distributions of classes which are obtained by plural class taps and allocates them to the converted classes in the order of higher frequencies. Thus, a class conversion rule which converts the number of classes obtained by sorting classes into a smaller number of classes is generated. A learning part 20 calculates an estimated coefficient serving as a data conversion rule for every class that is converted by the class conversion rule. A data conversion part 30 produces the pixel value of output data by linear primary connection secured between plural estimated taps and the estimated coefficient. The estimated coefficient is calculated for every converted class. A class whose sorting result is converted by the class conversion rule is also generated at the part 30.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image information conversion apparatus and method suitable for use in, for example, a television receiver.

[0002]

2. Description of the Related Art There has been known an image processing apparatus which linearly interpolates an input image signal to double the number of pixels (the number of lines) in a vertical direction. Such an image processing apparatus can be applied to, for example, converting an interlace system to a progressive system. This conversion is performed to reduce line flicker caused by the interlace system. For example, in a graphics image, there is a problem that line flicker is conspicuous. When displaying a graphics image, the progressive system can achieve higher image quality than the interlace system.

A conventional image information conversion apparatus of this type performs a motion determination process on an input 525i signal (interlaced signal of 525 scanning lines), and performs inter-frame interpolation when there is no motion. If there is motion, intra-field interpolation is performed. The inter-field interpolation forms a signal of a new line using the signal of the line of the previous field, and the intra-field interpolation forms a signal of the new line by averaging the signals of the upper and lower lines of the same field.

The conventional image information conversion device merely performs vertical interpolation on the basis of an input image signal, so that the resolution does not become higher than that of the base SD signal.
Also, the line created with the average value has a problem that the difference in resolution between the existing line and the interpolation line is noticeable because the vertical resolution is deteriorated as compared with the existing line. Furthermore, when there is noise in the image signal, and when the average value of the upper and lower lines is used, random noise is added, and noise is reduced in the created line.
As a result, the created line in which the noise has been reduced and the existing line in which the noise has been reduced appear alternately, and the image quality deteriorates. Further, there is a problem that when the interpolation method (still image processing and moving image processing) is switched based on the result of the motion detection, the image quality is greatly deteriorated when the motion detection is erroneous.

[0005] In addition to increasing the spatial resolution,
The resolution in the time direction may be increased. For example, a field doubling process of setting the field frequency to twice the original image signal may be performed to reduce field flicker. In the conventional field doubling process, the same field is read twice using a field memory. Therefore, even if flicker can be reduced, the resolution in the time direction cannot be increased.

In order to solve such a problem of the conventional image information conversion apparatus, the present applicant has already proposed an image information conversion apparatus based on a classification adaptive processing. Hereinafter, the previously proposed image information conversion device will be described.

[0007] The class adaptation process is different from the process of forming a high resolution signal by the conventional interpolation process. That is, the class classification adaptive processing has storage means for performing class division according to a three-dimensional (spatio-temporal) distribution of a video signal level as an input signal and storing a prediction coefficient value obtained by learning in advance for each class. This is a method of outputting an optimum estimated value by an operation based on a prediction formula, and the resolution can be increased to be higher than that of the input video signal by the class classification adaptive processing.

FIG. 16 schematically shows the entire configuration of a conventional image information conversion apparatus. In FIG. 16, 100 is a learning unit, and 200 is a data conversion unit. Learning unit 10
0 receives the student data and the teacher data, and calculates a prediction coefficient as a data conversion rule. The prediction coefficient and the input data are input to the data conversion unit 200, and output data after conversion is obtained from the data conversion unit 200. For example, when performing image information conversion in which the number of lines in the vertical direction is doubled, the teacher data is image data having twice the number of lines, the student data is the same image as the teacher data, and the number of lines is one. Image data.

FIG. 17 shows an overall flow of the processing. First, learning is performed for each result of the class classification, and a data conversion rule is created (step S100). This is learning unit 1
This is a process performed by 00. Next, data conversion is performed using a data conversion rule corresponding to the result of the classification (step S200). This is because the data conversion unit 200
This is the process performed by

FIG. 18 shows the configuration of the learning section 100. The student data is supplied to the class classification circuit 101, and a class is determined according to the level distribution of a plurality of pixel values (class taps) of the student data. The class is supplied to the learning circuit 102. Student data and teacher data are supplied to the learning circuit 102, and the prediction coefficient calculation circuit 103
A prediction coefficient is determined. The learning circuit 102 and the prediction coefficient calculation circuit 103 obtain prediction coefficients by the least square method. In this case, a prediction coefficient is calculated for each class.

FIG. 19 shows the flow of processing of the learning section 100. First, student data and teacher data are input (step S101). The student data is classified into classes (step S102). Learning is performed for each result of the class classification (step S103). Since the least squares method is used, the pixel value of the prediction tap and the true value are added to the normalization matrix.
A sufficient amount of data is used so that the learning is general. It is determined in step S104 whether or not the student data and the teacher data have been completed. If the data has been completed, in step S105, a prediction coefficient is calculated for each class classification result from the learning result (normalized matrix data) using a matrix solution. calculate.

FIG. 20 shows the configuration of the data converter 200. The input data is supplied to a class classification circuit 201 and a data conversion circuit 202. The classification circuit 201
Classification is performed based on the level distribution of the class taps of the input data. The result of the class classification is the data conversion circuit 20
2 is supplied. The data conversion circuit 202 creates a pixel value of the output data using the prediction coefficient corresponding to the class and a plurality of pixel values (prediction taps) of the input data. That is, by a linear linear combination of the prediction coefficient and the prediction tap,
Create output pixel values. The prediction coefficient is obtained by learning, and is actually stored in a memory.

FIG. 21 shows a flow of processing of the data conversion section 200. First, input data is input (step S2).
01). Next, the input data is classified into classes (step S202). Then, data is created by a linear combination of a prediction coefficient corresponding to the result of the class classification and a prediction tap, which are obtained in advance by learning (step S20).
3). Output the created data (Step S20)
4).

[0014]

In the image information conversion apparatus and method using the above-described classification adaptive processing, it is effective to increase the number of classes in order to improve the conversion accuracy. However, if you increase the number of classes,
The number of prediction coefficients increases. This causes a problem that the data conversion processing time increases and the capacity of the memory for storing the prediction coefficients increases.

Accordingly, an object of the present invention is to provide an image information conversion apparatus and method capable of performing high-accuracy conversion without increasing the number of classes.

[0016]

In order to achieve the above-mentioned object, the invention according to claim 1 provides a method for connecting teacher data to student data having the same image as the teacher data and having a smaller number of pixels than the teacher data. In an image information conversion rule learning device that learns a data conversion rule, class information is formed based on a plurality of first pixels of student data located around a predetermined pixel position of teacher data, and is acquired in advance. A class conversion unit that outputs the converted class by converting the class information according to the class conversion rule, and a plurality of second data of the student data located around a predetermined pixel position of the teacher data for each converted class. Learning for learning a prediction coefficient that minimizes an error when a true value at a predetermined pixel position of teacher data is predicted by linear linear combination of a pixel and a plurality of prediction coefficients. It consists of a, class conversion rules,
This is a table for converting N class information into M converted classes, and is an image information conversion rule learning device that is acquired in advance.

According to a second aspect of the present invention, there is provided an image information converting apparatus for generating a pixel which does not exist in an input image signal, wherein a plurality of first image signals of the input image signal located around a pixel position to be generated are provided. Based on the pixels, the class information is formed, and the class information is converted according to a class conversion rule acquired in advance, thereby outputting a converted class. A plurality of second pixels of the input image signal located;
A prediction coefficient obtained in advance for each conversion class is input, and a data conversion unit that generates a pixel by a linear estimation expression of the second pixel and the prediction coefficient is included. It is a table for converting information into M converted classes, and is an image information conversion device that is acquired in advance.

According to a fourth aspect of the present invention, there is provided an image information conversion rule learning method for learning a data conversion rule between teacher data and student data having the same image as the teacher data and having a smaller number of pixels than the teacher data. The class information is formed based on a plurality of first pixels of the student data located around a predetermined pixel position of the data, and the class information is converted according to a previously obtained class conversion rule. By performing a class conversion step of outputting a class and, for each converted class, a linear linear combination of a plurality of second pixels of the student data located around a predetermined pixel position of the teacher data and a plurality of prediction coefficients, A learning step for learning a prediction coefficient that minimizes an error when predicting a true value of a predetermined pixel position of the teacher data;
The class conversion rule is a table for converting N class information into M converted classes, and is an image information conversion rule learning method characterized by being acquired in advance.

According to a fifth aspect of the present invention, in the image information conversion method for generating a pixel not existing in the input image signal, a plurality of first image signals of the input image signal located around the pixel position to be generated are provided. Based on the pixels, form the class information, according to a class conversion rule acquired in advance, by converting the class information, the step of class conversion to output a converted class, after the conversion class,
A plurality of second pixels of the input image signal located around the pixel position and a prediction coefficient obtained in advance for each conversion class are input, and a pixel is estimated by a linear estimation expression of the second pixel and the prediction coefficient. The class conversion rule is a table for converting the N class information into the M converted classes, and is obtained in advance by the image information conversion method. is there.

According to the present invention, the class information is converted into a smaller number of classes in accordance with a class conversion rule created in advance, and learning and data conversion are performed using the converted class. Therefore, even if the number of classes is reduced, the accuracy of class classification can be improved.

[0021]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below. FIG. 1 schematically shows the entire configuration of an image information conversion device according to the present invention. In FIG. 1, 10 is a class conversion rule creation unit, 20 is a learning unit,
30 is a data conversion unit. FIG. 2 shows the overall flow of the processing of one embodiment. First, a class conversion rule is created from some input data (step S10). This is a process performed by the class conversion rule creating unit 10. Next, class classification is performed, and learning is performed for each result obtained by converting the result of the class classification according to the class conversion rule.
A prediction coefficient is calculated (Step S20). This is a process performed by the learning unit 20. Next, data conversion is performed using a prediction coefficient corresponding to a result obtained by converting the result of the class classification according to the class conversion rule (step S3).
0). This is a process performed by the data conversion unit 30. Note that the description of both the configuration shown in the block diagram and the processing shown in the flowchart means that it can be realized by either hardware or software. The same applies to the following description.

FIG. 3 shows the configuration of the class conversion rule creating unit 10. The student data is supplied to the class classification circuit 11, and the class is determined according to the level distribution of a plurality of pixel values (class taps) of the student data. The class is supplied to the class conversion rule creation circuit 12, and a class conversion rule is created.

FIG. 4 shows the flow of the process of the class conversion rule creating section 10. First, input data is input (step S11). Next, the input data is classified into classes (step S12). The result of the class classification is stored (step S13). It is checked in step S14 whether the input data has been completed. If the input data has not been completed, the process returns to step S11 and returns to steps S12 and S1.
3. Classification processing is performed in S14. When the input data is completed, a class conversion rule is created from the result of the class classification (step S15). The class conversion rule is a rule for converting the result of the classification of the input data into another class. By performing the class conversion, the number of classes can be reduced without lowering the accuracy.

FIG. 5 shows the configuration of the learning section 20. The student data is supplied to the class classification circuit 21, and the class is determined according to the level distribution of a plurality of pixel values (class taps) of the student data. Class classification result is class conversion circuit 2
2 is supplied. In the class conversion circuit 22, the classification result is converted into a smaller number of classification results according to the class conversion rule. The converted class classification result is called a converted class. By the class conversion circuit 22,
For example, the result of the class classification circuit 21 having 17 bits is converted into 11 bits by the class conversion circuit 22.

The student data, the teacher data and the converted class are supplied to the learning circuit 23. The prediction coefficient is calculated by the learning circuit 23 and the prediction coefficient calculation circuit 24. The learning circuit 23 and the prediction coefficient calculation circuit 24 obtain prediction coefficients by the least square method. In this case, a prediction coefficient is determined for each class after conversion.

FIG. 6 shows a flow of processing of the learning section 20.
First, student data and teacher data are input (step S21). Classify student data (step S
22). The class is converted using the class conversion rule (step S23). Learning is performed for each class after conversion (step S24). Since the least squares method is used, the pixel value of the prediction tap and the true value are added to the normalization matrix. A sufficient amount of data is used and learned to obtain a predictive coefficient with generality. Whether or not the student data and the teacher data have been completed is determined in step S25. If not, the process returns to step S21, and the processes of steps S22 to S25 are repeated. Step S25
In step S26, when the student data and the teacher data are completed, in step S26, a prediction coefficient is calculated for each class after conversion by using a matrix solution from the learning result (normalized matrix data).

FIG. 7 shows the configuration of the data conversion unit 30.
The input data is supplied to the class classification circuit 31 and the data conversion circuit 33. The class classification circuit 31 performs class classification based on the level distribution of the class taps of the input data. The result of the class classification is supplied to the class conversion circuit 32. The class conversion circuit 32 converts a class according to a class conversion rule obtained by learning. Classification circuit 31
The class conversion circuit 32 is the same as the class classification circuit 21 and the class conversion circuit 22 in the learning unit 20. That is, the classification result is converted into a smaller number of classification results according to the class conversion rule (class conversion table).

The converted class from the class conversion circuit 32 is supplied to the data conversion circuit 33. Data conversion circuit 33
Creates an output pixel value by a linear linear combination of a prediction coefficient corresponding to the converted class and a plurality of pixel values (prediction taps) of input data. The prediction coefficient is obtained by learning, and is actually stored in a memory.

FIG. 8 shows the flow of processing of the data conversion unit 30. First, input data is input (step S3).
1). Next, the input data is classified into classes (step S
32). The obtained class classification result is converted using the class conversion rule (step S33). Predicted coefficients (data conversion rules) corresponding to the converted classes are obtained in advance by learning. Data is created by a linear combination of the prediction coefficients and the prediction taps (step S34). Output the created data (step S3
5).

The above-described embodiment of the present invention will be described in more detail. FIG. 9 shows the class conversion rule creation unit 1.
0 shows a more detailed configuration. The input data is supplied to the class tap selection circuit 11a. As input data, student data used for learning is used. However, other than the student data, data for creating a class conversion rule may be used. Further, different types of images may be used as input data, and a change rule may be created for each type of different images. In this case, since a plurality of class conversion tables are created, a switching signal for switching the table according to the type of image is supplied to the class conversion circuit 22. The types of images are animation, natural images, computer graphics, still images, and the like.

The class tap selection circuit 11a selects, for example, 12 pixels of the input data indicated by black circles when creating a pixel value at a position indicated by X in FIG. FIG. 10 shows partial images included in the current frame and the frames before and after the current frame and at the same spatial position. Each frame is composed of an even field and an odd field. The field frequency is 60 Hz and the frame frequency is 30 Hz. FIG. 10 shows only the line corresponding to the line position, the class tap (black circle pixel), and the prediction tap (black circle pixel and white circle pixel).
Illustration of pixels not used as taps is omitted.

By setting the class taps and the prediction taps over a plurality of fields, in addition to the spatial level distribution, the pixel data from these taps has a component of a temporal image change (ie, motion). become. Therefore, the value of the created pixel reflects the motion. A space class detection circuit that detects a spatial class corresponding to the level distribution pattern and a motion class detection circuit that detects a motion class corresponding to the amount of motion are provided separately, and the space class and the motion class are combined. Class may be formed.

The pixel value of the class tap selected by the class tap selection circuit 11a is 1 bit ADRC (Adapti).
ve Dynamic Range Coding) circuit 11b to convert each pixel value of, for example, 8 bits into 1 bit. This process is performed to prevent the number of classes from becoming enormous. As information compression means, other than ADRC, DPCM (prediction coding), VQ (vector quantization)
Or other compression means.

Originally, the ADRC is a VTR (Video Tape R)
ecoder) is an adaptive requantization method developed for high-efficiency coding. However, since a local pattern of a signal level can be efficiently represented by a short word length, in this example, ADRC is used.
Is used to generate codes for spatial class classification. The 1-bit ADRC defines the dynamic range of the class tap as D
R, the data level of the pixel of the class tap is L, and the requantization code is Q (0 or 1).
According to (2), the distance between the maximum value MAX and the minimum value MIN is 2
Divide and requantize.

DR = MAX−MIN + 1 (1) Q = {(L−MIN + 0.5) × 2 / DR} (2) where {} means truncation processing.

The 1-bit ADRC circuit 11b generates an output having a bit number equal to the number of class taps (this is called an ADRC class). The ADRC class is supplied to the class frequency data creation circuit 12a. The class frequency data creation circuit 12a counts the frequency of occurrence for each ADRC class number. The output of the class frequency data creation circuit 12a is supplied to the conversion table creation circuit 12b. The conversion table creation circuit 12b rearranges the ADRC classes in descending order of frequency, and newly assigns a class number (hereinafter referred to as a frequency class) to the rearranged ADRC classes. The frequency class corresponds to the converted class. The conversion table creating circuit 12b creates a conversion table (conversion rule) indicating the correspondence between the ADRC class and the frequency class.

FIG. 11A is an example of the appearance frequency of the ADRC class. As an example, an image of 50 frames is used as input data. It is preferable that the input data include as many types of images as possible. A class tap is set for the image of each frame. FIG. 10 shows an example in which a class tap of 12 pixels is set.
FIG. 11A is an example of a class tap of 17 pixels. That is, as the ADRC class, 2 ¹⁷ (= 131,07)
There are 2) ways. In FIG. 11A, the horizontal axis represents the ADRC class, and the vertical axis represents the frequency normalized value (%). The normalization is performed by dividing the appearance frequency of each ADRC class by (the total number of frames × the number of pixels (per frame)) of the input data.

By rearranging the appearance frequencies of the ADRC classes in descending order, the frequency classes shown in FIG. 11B are obtained. The horizontal axis in FIG. 11B is a frequency class. The frequency class is, for example, 11 bits and can represent 2048 classes. 2047 ADRs from the most frequent ADRC class in the most frequent order
Frequency classes 0 to 2046 are sequentially assigned to the C class. For the 2048th and subsequent ADRC classes, 2047 frequency classes are collectively assigned. In this way, 131,072 ADRC classes are converted into 2048 frequency classes. Since the number of classes is reduced with reference to the frequency, it is possible to prevent a decrease in accuracy even with a small number of classes.

FIG. 12 shows a conversion table.
This shows how the N class of the DRC class is converted to the M (<N) class of the frequency class. After the data for creating the class conversion table is input, the M−1 frequently used ADRC classes are associated with the frequency classes on a one-to-one basis, and the remaining (N− (M−1)) ADRC classes are used. The classes are collectively assigned to one frequency class. In this way, a class conversion table from N to M is created.

FIG. 13 shows a specific example of the class conversion table. The ADRC class number represented by 17 bits is converted into a frequency class represented by 11 bits. In FIG. 13, it is assumed that the frequency class number = 2047 is collectively assigned to other ADRC classes. Such a class conversion table is commonly used in the learning unit and the data conversion unit.

FIG. 14 shows a more detailed configuration of the learning section 20. The student data is supplied to the class tap selection circuit 21a. The class tap selection circuit 21a is the same as that shown in FIG.
When creating a pixel value at a position indicated by x in the above, 12 pixels of student data indicated by black circles around the pixel value are selected.
The pixel value of the class tap selected by the class tap selection circuit 21a is supplied to the 1-bit ADRC circuit 21b, and each 8-bit pixel value is converted to 1 bit.

The 1-bit ADRC circuit 21b generates an ADRC class having the number of bits equal to the number of class taps. The ADRC class is supplied to a class conversion circuit 22 that performs class conversion. The class conversion circuit 22 converts the conversion table (class conversion rule) created in advance as described above.
And convert the class according to
A converted class (frequency class) is generated.

Frequency class is normalized matrix selection circuit 2
3a. The normalization matrix from the normalization matrix selection circuit 23a is supplied to the normalization matrix addition circuit 23b. Normalization matrix addition circuit 2
The prediction tap (student data) from the prediction tap selection circuit 23c and the true value (teacher data) from the true value selection circuit 23d are supplied to 3b. The prediction tap is shown in FIG.
As shown in FIG. 7, there are 20 pixels set over three temporally continuous fields. The output of the normalization matrix addition circuit 23b is supplied to the prediction coefficient calculation circuit 24. The prediction coefficient calculation circuit 24 is a circuit that calculates a prediction coefficient by solving a normalization matrix.

The learning performed by the learning unit 20 is such that, for each frequency class, which is a converted class, the sum of squares of the error between the predicted value obtained by the linear linear combination of the prediction tap and the prediction coefficient and the true value is minimized. This is a process for obtaining a prediction coefficient that becomes The learning process in the case of performing prediction using n pixels will be described below.

When the levels of the input pixels selected as the prediction taps are x ₁ , ‥‥, and x _n , and the output pixel level is y, the prediction coefficients w ₁ , ‥‥,
Set an n-tap linear estimation equation with w _n . This is shown in the following equation (3). Learning ago, w _i is undetermined coefficients.

Y = w ₁ x ₁ + w ₂ x ₂ + ‥‥ + w _n x _n (3) Learning is performed on a plurality of signal data for each frequency class. When the number of data is m, the following equation (4) is set according to the equation (3).

Y _k = w ₁ × _{k 1} + w ₂ × _{k 2} + ‥‥ + w _n × _kn (4) (k = 1,2, ‥‥ m) When m> n, the prediction coefficients w _i and ‥‥ w _n Is not uniquely determined, the element of the error vector e is defined by the following equation (5), and a prediction coefficient that minimizes the equation (6) is obtained. This is a so-called least squares solution.

E _k = y _k- {w ₁ x _k1 + w ₂ x _k2 + ‥‥ + w _{n xkn} } (5) (k = 1, 2, ‥‥ m)

[0049]

(Equation 1)

Here, the partial differential coefficient based on w _{i in} equation (6) is obtained. What is necessary is just to find each coefficient w _i so that the following equation (7) is set to `0`.

[0051]

(Equation 2)

Hereinafter, X _ij , Y as shown in equations (8) and (9)
_{When i} is defined, equation (7) is obtained by using equation (10) using a matrix.
Is rewritten to

[0053]

(Equation 3)

[0054]

(Equation 4)

[0055]

(Equation 5)

This matrix is called a normalization matrix. The normalization matrix adding circuit 23b in FIG. 14 adds a normalization matrix for each frequency class.

After the input of data of a sufficient number of frames for learning is completed, the normalization matrix adding circuit 23b
The normalized matrix data is output to the prediction coefficient calculation unit 24. The prediction coefficient calculation unit 24 solves w _i using a general matrix solution such as a sweeping-out method of the normalization matrix, and calculates a prediction coefficient. The prediction coefficient calculation unit 24 writes the calculated prediction coefficient into a prediction coefficient memory, not shown.

As a result of learning as described above, the prediction coefficient memory for estimating the target pixel y for each frequency class, which can be statistically estimated to be closest to the true value, is stored in the prediction coefficient memory. . The prediction coefficients stored in the prediction coefficient memory are used in the data conversion unit 30.

FIG. 15 shows a more detailed configuration of the data conversion unit 30. Input data is a class tap selection circuit 31a
Supplied to The class tap selection circuit 31a generates a pixel value at a position indicated by x in FIG.
Twelve pixels of student data indicated by black circles around the pixel are selected. The pixel value of the class tap selected by the class tap selection circuit 31a is supplied to the 1-bit ADRC circuit 31b, and each 8-bit pixel value is converted to 1 bit.

The 1-bit ADRC circuit 31b generates an ADRC class having the number of bits equal to the number of class taps. The ADRC class is supplied to the class conversion circuit 32. The class conversion circuit 32 converts the ADRC class into the frequency class according to the conversion table created in advance as described above, and a converted class (frequency class) is generated at the output.

The frequency class from the class conversion circuit 32 is supplied to the prediction coefficient selection circuit 33a. The prediction coefficient selection circuit 33a includes a prediction coefficient memory in which prediction coefficients obtained by learning are stored for each frequency class. The prediction coefficient selected by the prediction coefficient selection circuit 33a is the product-sum operation circuit 3
3b. The prediction tap selected by the prediction tap selection circuit 33c is supplied to the product-sum operation circuit 33b.
An output pixel value is created by a linear linear combination of a prediction coefficient and a prediction tap.

Although not shown, when performing image conversion such that the number of lines is twice as large as the input data, the output data is processed at a line double speed by a line memory, that is, a line whose line frequency is doubled. Receive double-speed processing. In this case, conversion processing for doubling the number of pixels in the horizontal direction may be performed simultaneously. Further, the output video signal subjected to the line double speed processing is supplied to a CRT display. The synchronization system of the CRT display is configured so that an output video signal can be displayed. The input video signal is a broadcast signal or VT
A playback signal of a playback device such as R is supplied. That is, this example can be incorporated in a television receiver.

The present invention is applicable not only to the conversion of image information for doubling the number of lines and / or the number of horizontal pixels but also to the case of obtaining a field doubled output image for doubling the field frequency. Applicable. In that case,
A field memory is connected to the product-sum operation circuit 33b, and the field frequency is doubled. Furthermore, the present invention can be applied to a case where an output image signal having twice the number of vertical pixels and twice the field frequency is formed.

[0064]

According to the present invention, when the process of increasing the number of pixels in space and / or in the time direction is performed by the class classification adaptive process, a class conversion rule is created in advance, and the class information is reduced according to the class conversion rule. Has been converted to the class. Therefore, the number of data conversion rules (prediction coefficients) can be reduced, the processing time for learning and data conversion can be reduced, and the capacity of the memory for storing the data conversion rules can be reduced. In addition, since class conversion rules are created that match the input data trends,
Even if the number of classes is reduced, deterioration of accuracy can be prevented.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an overall configuration of an embodiment of the present invention.

FIG. 2 is a flowchart showing an overall process according to an embodiment of the present invention.

FIG. 3 is a block diagram illustrating an example of a class conversion rule creating unit according to the embodiment of the present invention;

FIG. 4 is a flowchart of an example of a class conversion rule creation process according to an embodiment of the present invention.

FIG. 5 is a block diagram illustrating an example of a learning unit according to the embodiment of the present invention;

FIG. 6 is a flowchart illustrating an example of a learning process according to an embodiment of the present invention.

FIG. 7 is a block diagram illustrating an example of a data conversion unit according to an embodiment of the present invention.

FIG. 8 is a flowchart illustrating an example of a data conversion process according to an embodiment of the present invention.

FIG. 9 is a block diagram illustrating a more detailed configuration of a class conversion rule creation unit according to the embodiment of the present invention.

FIG. 10 is a schematic diagram illustrating an example of a class tap and a prediction tap according to the embodiment of the present invention.

FIG. 11 is a schematic diagram illustrating a frequency distribution of ADRC classes used for describing a class conversion rule creation process according to an embodiment of the present invention.

FIG. 12 is a schematic diagram used for describing a class conversion process according to an embodiment of the present invention.

FIG. 13 is a schematic diagram illustrating an example of a class conversion table according to an embodiment of the present invention.

FIG. 14 is a block diagram illustrating a more detailed configuration of a learning unit according to the embodiment of the present invention.

FIG. 15 is a block diagram illustrating a more detailed configuration of a data conversion unit according to an embodiment of the present invention.

FIG. 16 is a block diagram showing an overall configuration of a conventional image information conversion device.

FIG. 17 is a flowchart showing the overall processing of a conventional image information conversion device.

FIG. 18 is a block diagram illustrating an example of a learning unit in a conventional image information conversion device.

FIG. 19 is a flowchart of an example of a learning process in a conventional image information conversion device.

FIG. 20 is a block diagram illustrating an example of a data conversion unit in a conventional image information conversion device.

FIG. 21 is a flowchart of an example of a data conversion process in a conventional image information conversion device.

[Explanation of symbols]

10: Class conversion rule creation unit, 20: Learning unit, 30: Data conversion unit, 11: Class classification unit, 11a: Class tap selection circuit, 11b ...
1-bit ADRC circuit, 12: Class conversion rule creation unit, 12a: Class frequency data creation circuit, 12b
... Conversion table creation circuit

 ──────────────────────────────────────────────────続 き Continued on the front page F term (reference) 5C059 KK00 LA00 LB15 LB16 PP01 PP04 PP12 SS21 TA08 TB04 TB05 TC00 TD10 5C063 BA03 CA05 CA16 CA40

Claims (6)

[Claims] An image information conversion rule learning apparatus for learning a data conversion rule between teacher data and student data having the same image as the teacher data and having a smaller number of pixels than the teacher data. Class information is formed based on a plurality of first pixels of the student data located around the position, and the class information is converted according to a class conversion rule acquired in advance, thereby outputting a converted class. A class conversion unit, and for each of the converted classes, a linear primary combination of a plurality of second pixels of the student data and a plurality of prediction coefficients located around a predetermined pixel position of the teacher data to generate teacher data. And a learning unit for learning a prediction coefficient for minimizing an error when a true value of a predetermined pixel position is predicted. The class conversion rule includes N classes. An image information conversion rule learning device, which is a table for converting information into M converted classes, which is obtained in advance. 2. An image information conversion apparatus which generates a pixel which does not exist in an input image signal, based on a plurality of first pixels of the input image signal located around a pixel position to be generated. A class conversion unit that forms class information and converts the class information according to a class conversion rule acquired in advance, thereby outputting a converted class. The class conversion unit is located around the pixel position. A plurality of second pixels of the input image signal and a prediction coefficient obtained in advance for each of the conversion classes are input, and data for generating a pixel by a linear estimation formula of the second pixel and the prediction coefficient The class conversion rule is a table that converts N class information into M converted classes, and is obtained in advance. Information conversion device. 3. The class conversion rule according to claim 1, wherein the class conversion rule determines an appearance frequency of each of the class information formed based on the first pixel, and selects the class information in order from the one having the highest appearance frequency. Created by allocating the converted class to the M-1 pieces of the class information and allocating one converted class to the remaining N- (M-1) pieces of the class information. An apparatus characterized in that: 4. An image information conversion rule learning method for learning a data conversion rule between teacher data and student data having the same image as the teacher data and having a smaller number of pixels than the teacher data, comprising the steps of: Class information is formed based on a plurality of first pixels of the student data located around the position, and the class information is converted according to a class conversion rule acquired in advance, thereby outputting a converted class. A class conversion step, and, for each of the converted classes, a linear primary combination of a plurality of second pixels of the student data and a plurality of prediction coefficients located around a predetermined pixel position of the teacher data, whereby the teacher And a learning step for learning a prediction coefficient for minimizing an error when a true value of a predetermined pixel position of the data is predicted. Is a table for converting N class information into M converted classes, and is a method for learning an image information conversion rule, which is acquired in advance. 5. An image information conversion method for generating a pixel that does not exist in an input image signal, the method comprising: generating a pixel that does not exist in an input image signal based on a plurality of first pixels of the input image signal located around a pixel position to be generated. A class conversion step of forming class information and converting the class information according to a class conversion rule obtained in advance, thereby outputting a converted class; and A plurality of second pixels of the input image signal to be input and prediction coefficients obtained in advance for each of the conversion classes are input, and pixels are generated by a linear estimation expression of the second pixels and the prediction coefficients. And a data conversion step. The class conversion rule is a table for converting N class information into M converted classes, and is obtained in advance. An image information conversion method characterized by the following. 6. The class conversion rule according to claim 4, wherein the class conversion rule obtains an appearance frequency of each of the class information formed based on the first pixel, and selects the class information in order from the one having the highest appearance frequency. Created by allocating the converted class to the M-1 pieces of the class information, and allocating one converted class to the remaining N- (M-1) pieces of the class information. A method characterized in that: