JPH0591532A - Image filter and adaptive type image filter learning method - Google Patents

Abstract

PURPOSE:To output a signal in an optimum state by collectively judging the features of an input image and optimizing a circuit state. CONSTITUTION:In a filter provided with an input terminal 11, an output terminal 12 and a coefficient variable means 18, plural feature extracting means 13 to 15 respectively extract the features of an input signal. A control signal generating means 16 generates a control signal from respective output signals of the means 13 to 15 based upon fuzzy inference. The means 18 is controlled based upon the control signal and a circuit state for the input is optimized in accordance with the judged result of the composite features of the input image.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image filter, and more particularly to an image filter which adaptively changes a circuit constant according to an input image to improve the image quality.

[0002]

2. Description of the Related Art In order to improve the image quality of a television receiver or a video tape recorder (VTR), various image filters are used in each part in the video signal processing. For example, there is a Y / C separation circuit for separating the composite signal into a luminance signal and a chrominance signal, and there is a luminance signal noise reduction circuit or a noise canceller for reducing the noise contained in the luminance signal, and the noise contained in the chrominance signal. There is a color signal noise reduction circuit to reduce noise, and an edge enhancement circuit to obtain a clear image.

These image filters have been individually improved to the desired characteristics, and in fact television receivers and V receivers.
The image quality of TR has also improved against the background of these technologies. For example, regarding a luminance signal noise reduction circuit, a low-pass filter type one-dimensional noise reduction circuit that reduces noise (high frequency noise) on the horizontal axis of the luminance signal, or a comb that reduces noise using the vertical axis. A two-dimensional filter type two-dimensional noise reduction circuit and a three-dimensional noise reduction circuit that reduces noise using the time axis have been realized. Also, a noise reduction circuit that detects the correlation between input signals and switches the operation mode according to the input image state has been put into practical use.

For example, in the case of the motion adaptive three-dimensional Y / C separation circuit, in the case of a still image, the three-dimensional Y / C separation is performed to give the maximum Y / C separation effect, and when the motion is detected, the screen is displayed. There is a method in which all or part of the above is switched to two-dimensional Y / C separation to improve blurring for moving images, which is a three-dimensional defect.

Similarly, in the case of 3-line Y / C separation,
Normally, 3-line Y / C separation, which has better Y / C separation performance than a 1-dimensional filter, is performed, and when a signal that is not suitable for 3-line Y / C separation comes, it is detected and 1-dimensional filter separation or color There is one that suppresses the defect of 3-line Y / C separation by suppressing the output.

As for the luminance signal noise reduction circuit, since the noise reduction circuit causes a reduction in noise and a deterioration in resolution at the same time, the noise reduction circuit normally operates to reduce the noise contained in the signal and reduce the noise. There is a method of detecting the amount of the non-correlation component of the signal for performing the above, and turning off the noise reduction circuit to prevent the deterioration of resolution when it is determined that there is no correlation. Similarly, there are color signal noise reduction circuits and contour enhancement circuits that switch the operation by detecting the amount of uncorrelated components.

[0007]

However, the degree of adaptation of the conventional image filter is not sufficient, and the image quality is not sufficiently improved. In the image filter, there is always a trade-off between the effect and the harmful effect, and the optimum point changes every moment depending on the input image, but the conventional image filter is not sufficiently adapted to it.

In the conventional method, an engineer devises a method, a procedure, and a circuit for discriminating the decorrelation to improve the image quality by half trial and error. Therefore, in a certain kind of image, an operation that cannot be considered by the engineer is performed. There was a case to show. That is, in the examined images, there is a problem that the decorrelation is effective, and even if the image quality is improved, it is harmful in the case of certain images that have not been examined. In addition, since the adaptation operation is immature, the effect of improving the image quality by image adaptation is small, and even if the person who is familiar with the image knows the effect of improving the image quality, there is a problem that it cannot be seen as an effect by the general public.

These problems are caused by imperfect adaptation operations. When this problem is subdivided, the insufficient adaptive operation is caused by two problems. The two points will be described with examples.

In the case of the luminance signal noise reduction circuit, noise reduction is performed by extracting a difference signal (non-correlation component) between the signal delayed in the horizontal scanning time and the original signal not delayed and subtracting this from the original signal. The signal state is detected by the magnitude of the difference signal. That is, when the difference signal is larger than a certain value, the subtraction operation is stopped to prevent deterioration of resolution.

The first problem is that the signal state detection operation is insufficient. It operates by detecting only the amount of uncorrelated components, but the trade-off between noise reduction effect and resolution degradation cannot be positioned at the optimum point by this alone. Obviously, not only the amount of uncorrelated components, but also the amount of noise contained in the signal, the spectrum of the signal, and the like are greatly affected. Also, the sensitivity of the human eye differs depending on the brightness, etc., and the sensitivity to noise differs in bright areas. That is, in order to more adapt to the image quality and improve the image quality, it is necessary to detect these states and comprehensively examine the image state. In the past, there was no method for comprehensively recognizing and processing the image quality state.

Secondly, although the image originally viewed by human beings is an analog amount, only digital control is possible to switch the operation depending on the presence or absence of correlation, and an intermediate signal comes. In this case, there is a problem that the adaptive operation is not performed. The state of the input image is infinite when viewed in detail, and control should be continuous. As for the noise reduction circuit, the optimum point of the trade-off is that some images exist where the noise reduction is maximized, and some images have the best noise reduction with a 10% effect. There is also an image called optimal. For example, suppose that there is a signal that the decorrelation detector can detect within the limit frame. At this time, the conventional noise reduction circuit detects the decorrelation, turns off the noise reduction circuit, and outputs a signal containing noise at a considerable ratio. In such a case, it is originally desirable that the noise reduction circuit operates so that the noise reduction ability is moderate.

An object of the present invention is to provide an adaptive image filter that comprehensively identifies the state of an input image and optimizes the state of the image filter according to the state. Further, it is another object of the present invention to provide an image filter which can be easily designed for adaptation and is less likely to cause an adverse effect even on an image which was not predicted at the time of design.

[0014]

In an image filter having an input terminal, an output terminal, and a coefficient varying means, a plurality of feature extracting means for extracting features of an input signal, and control from output signals of the feature extracting means And a control signal generating means for generating a signal, and the image filter is configured to control the coefficient varying means by the control signal of the control signal generating means. The control signal generating means is realized by means using fuzzy inference or neural network. If these control signal generating means are programmed in advance so as to comprehensively judge a plurality of features extracted from the input image and generate an appropriate control signal, the input signal (information ) Is comprehensively judged and a control signal of an appropriate level (or value) can be generated.

[0015]

Since the adaptive image filter constructed as described above comprehensively judges various characteristics, it is less likely that a detection error or malfunction will occur, and the adaptive image filter can be optimized to a state closer to the true optimum value than in the past. it can.

Since the control signal generating means using fuzzy inference can describe an optimization program by rules, it is possible to decompose and rearrange operations by each extracted feature, and thus to extract various types of features and accurately. Can be controlled. In addition, since analog (ambiguous) calculations can be performed, signals that do not have clear characteristics can be handled appropriately without malfunction, and with conventional discrete control represented by switching. Instead, control according to the degree becomes possible. Furthermore, since the method of giving control is decomposed into rules, it is easy to deal with changes in control, and the addition of new feature extraction means and rules allows smooth development to higher-level adaptation. Like That is, the drawbacks of the conventional adaptive image filter can be significantly reduced, and an image filter having characteristics closer to ideal can be realized.

The same effect can be obtained in the case of the control signal generating means using the neural network, but the optimization program is not described by the rule, but is performed by teaching the neural network. Therefore, when the control is changed. It is inferior to the control signal generating means of fuzzy reasoning in that the neural network has to be re-educated and this process has to be redone. However, this has the advantage that it can be realized by making the neural net experience learning even when the shape of the rule cannot be clarified.

[0018]

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

FIG. 1 shows an embodiment of the present invention. This adaptive image filter includes an input terminal 11 for inputting a signal and an output terminal 12 for outputting a signal, and three feature extracting means 13, 14, 15 are connected to the input terminal 11 and the feature extracting means 1 is provided.
The outputs of 3, 14, and 15 are supplied to the control signal generating means 16, and the output of the control signal generating means 16 is an adaptive image filter connected to the coefficient changing means 18 of the image filter 17.

Although the input image signal changes every moment, the feature extracting means 13 to 15 extract the feature at each moment. The features extracted by the respective feature extracting means 13 to 15 are different. For example, some extract the amount of high frequency components, some extract the amount of low frequency components, and some are separated by a certain time. Extract the correlation with the part. The control signal generating means 16 comprehensively judges the extracted characteristics and outputs an appropriate control signal. Fuzzy inference or inference by neural network is used for the comprehensive judgment. The control signal generated in this manner is supplied to the coefficient changing means 18 to change the coefficient of the coefficient changing means 18 so as to be suitable for the input image signal examined by the feature extraction so as to be suitable for the input image signal. The characteristics of the image filter 17 are changed.

For example, a noise reduction circuit will be described in more detail. In the case of a signal having a large input noise, the coefficient is changed so that the image filter has a large noise reduction effect, and the noise is reduced. Output as a legible image. On the contrary, when the input noise is small, the coefficient is varied so that the noise reduction effect of the image filter is reduced, or the coefficient is varied so as to have the edge enhancement effect, and a clear image is output. It can be applied in this way.

Fuzzy reasoning has been known for a long time, and its application has been remarkable in recent years, but it has not been applied to an image filter. It is difficult to apply fuzzy inference to the signal itself input to the image filter, but it can be applied to the control part as in the present invention. Since the image filter requires a huge amount of calculation, real-time display is not possible unless the calculation is performed by hardware. For example, when a television signal is digitally processed, it is sampled and processed at a frequency (about 14 MHz) four times as high as that of a normal color subcarrier. In order to output an image in real time, a lot of arithmetic processing must be performed within this period, which is a calculation amount that cannot be processed even by the latest processor. For example, if a multiprocessor configuration such as having a processor for each pixel or for each line enables calculation by software, a large number of processors are required (525 for each line), so the cost is low. Becomes extremely expensive,
Moreover, the size of the device becomes extremely large, so it cannot be used for consumer use. According to the present invention, the image processing itself is performed by hardware, and the fuzzy inference is used only for the control signal to be optimized. With this configuration, since the essential part of image calculation can be performed by hardware, the circuit can be integrated in a single-chip IC even with the current technology. Further, since the fuzzy inference part is a control signal, the amount of information and the amount of calculation can be greatly reduced from the original image signal. The number of bits of the control signal may be half to 1/4 that of the image signal. It also has a smaller number of points to calculate and is much less computationally complex.

The fuzzy reasoning program may be stored and calculated every time the detection signal is received by the processor, but a table lookup method is also possible. That is,
If the fuzzy inference is calculated in advance for all the data that may be input, the input data is used as the memory address, and the result of the fuzzy inference is stored in the data part of that, the feature extraction result will be obtained. The result of fuzzy inference (control signal) can be obtained simply by referring to the memory as the address. Unlike the image signal, the feature extraction signal may have a low resolution. Therefore, even if such a method is adopted, the memory amount does not become enormous and the amount can be easily stored in the IC.

For example, in FIG. 1, the feature extraction 1 is 4 bits (16 steps), the feature extraction 2 and the feature extraction 3 are 2 bits (4 steps), and the output control signal is 4 bits (1 step).
If there are 6 stages), the required address is 8 bits, and the control signal generating means can be realized by a 1024 bit memory. Moreover, since this can be a ROM, it can be integrated on the same chip as a very small part in the IC.

For comparison, consider that all the image filters are of the table look type. Then, since the input signal and the output signal, as well as the signal internally delayed, must be processed with 8 bits or more, a huge memory is required, which is not realistic. Considering the case of inputting three points, the address becomes 24 bits, and the required memory becomes huge as 1,341,728 bits. An increase of 1 bit in the address means that twice the memory is required, and the effect of reducing the number of bits by configuring only the control signal is extremely large.

Since the method of the present invention is performed by feature extraction, although the number of bits is small, the required memory amount suddenly increases as the number of features to be extracted increases. In this case, the memory amount can be reduced by using the control signal generating means having a multi-stage configuration. That is, the extracted features are classified into several groups, an intermediate control signal is generated for each group, and the intermediate control signal is further combined in the next stage to generate a control signal. By doing so, even if the number of features to be extracted is increased, the amount of memory can be suppressed to be small. For example, when handling 4 types of 4 bit features, the address is 16 bits and the memory is 262144.
Bits are required, but if these are divided into two groups of two types, the address of each group is 8 bits and the memory can be 1024 bits per group. A part for producing the control signal by synthesizing the intermediate control signal is also required, but it can be realized with a memory of 3072 bits in total, which is sufficiently practical. In the case of the present invention, such a method is possible because only the control signal is used for the table lookup.

If the essential part of the image filter 17 is prepared by hardware as in the present invention and only the control section generates a control signal by fuzzy inference, a compact yet sophisticated adaptive image filter can be realized. In addition, although the basic part of the image filter is promising technology, as for the adaptation part,
This is a future technology. Since the present invention realizes the adaptation unit by software, it is possible to improve the characteristics by the adaptation improvement by changing the program without changing the hardware part of the image quality filter. In addition, although it is a similar technology such as a television receiver, a laser disk, a VTR, etc., it has been difficult to use the conventional image filter for other devices because of its slightly different characteristics. The image filter of the invention has the advantage that the characteristics can be tuned only by changing the software, that is, rewriting the ROM, and the application range is wide. FIG. 2 ... Embodiment applied to vertical noise reduction circuit A more concrete embodiment is shown in FIG. This example is a luminance signal noise noise reduction circuit for television signals.

Basically, the input terminal 21 and the output terminal 22
Therefore, the input signal is input to the 1H delay means 23 and the subtraction means 24. The subtraction means 24 extracts the difference (non-correlation component) between the input signal and the signal delayed by the horizontal scanning period (1H) and supplies it to the limiter 25. The limiter 25 is a circuit that cuts off an uncorrelated component having an amplitude equal to or larger than the amplitude, subtracts the uncorrelated component from the input signal by the subtracting unit 26, reduces noise, and outputs the circuit. The transfer characteristic of this system is that when the amount of uncorrelated components does not reach the limiter,

Gain = 1.0-g (1.0-exp (-jωt) (1) where j is sqrt (-1), ω is the angular frequency, t is the delay time of the delay means, and g is variable. The amplification factor of the gain amplifier If g = 0.5, the noise reduction circuit has a gain of 1.0 for correlated signals and 0.0 for uncorrelated signals.

In the embodiment of the present invention, the detecting means 27 and 28 are connected to the input terminal 21 and the output terminal of the subtracting means 24, respectively, and their outputs are inputted to the control signal generating means 29. Then, the control signal output from the control signal generating means 29 controls the amplification factor of the variable gain amplifier 30. The variable gain amplifier 30 controls the gain of the output of the limiter 25 and supplies it to the subtracting means 26.

Although the noise reduction circuit reduces noise to form an image with less noise, it also has the disadvantage of deteriorating the resolution. Since the gain of the decorrelation signal is 0.0 in Eq. (1), the decorrelation signal is not output and is removed by the noise reduction circuit. Image signals often have vertical correlation, but many signals do not have correlation. In a vertical noise reduction circuit like this example, a non-correlation component is included except for a perfectly vertical line or a perfectly flat surface, but it is cut off. That is, the resolution deteriorates. I only need to remove the noise,
This is because noise and minute signals cannot be distinguished. Therefore, if the amount of the non-correlation component is larger than a certain amount, it is regarded as a signal and the noise reduction circuit is turned off (the amplification factor of the variable gain amplifier 30 is set to zero) to prevent the deterioration of resolution, and It is a conventional noise reduction circuit that if the amount is smaller than a certain amount, it is regarded as noise and removed.

Since there are various kinds of input image signals, it is not sufficient to control only by detecting the amount of the non-correlation component, and if the control is performed more finely, it is possible to suppress the adverse effect and reduce the noise. Therefore, the image quality can be improved. For example, human vision has a property that sensitivity of a bright image portion is low. Therefore, noise is less noticeable in bright image areas,
When the brightness is below the middle level, noise is likely to be noticed. If a noise reduction circuit that can adapt to such a property is realized, the image quality can be improved, and FIG. 2 shows an example in which it is realized.

The detection means 28 detects the amount of the uncorrelated component,
The detecting means 27 detects the brightness. Control signal generating means 2
9 generates a control signal by fuzzy inference from the two detection signals, and controls the amplification factor of the variable gain amplifier 30. In fuzzy reasoning, the following rules should be applied.

Rule ... (If) ... (Then) Rule 1 ... If the amount of the uncorrelated component is large ... Reduce the amplification factor of the variable gain amplifier Rule 2 ... If the amount of the uncorrelated component is small ... Variable gain Increase the amplification factor of the amplifier Rule 3 ... If the non-correlation component occurs frequently ... Reduce the amplification factor of the variable gain amplifier Rule 4 ... If the non-correlation component does not occur frequently ... Increase the amplification factor of the variable gain amplifier Rule 5 ... If it is bright (the brightness is high) ... Decrease the gain of the variable gain amplifier Rule 6 ... If it is dark (the brightness is low) ... Increase the gain of the variable gain amplifier (if) If fuzzy inference is performed using the rule of (then), the amplification factor of the variable gain amplifier 30 becomes maximum (0.5 in the case of FIG. 2) when it is dark and the non-correlation component is small,
When the brightness is large and the amount of the non-correlation component is large, the amplification factor of the variable gain amplifier 30 is minimized, and in the middle thereof, the control signal becomes a value derived by fuzzy inference.

FIG. 3 (A) ... Example of frequency characteristic for explaining the system of FIG. 2 FIG. 3 shows an example of change in frequency characteristic. In order to show that the valley portion of the frequency characteristic is changing, three types of characteristic examples are overlaid. When it is dark and the non-correlation component is small, the valley has the deepest characteristic in the figure, and the noise reduction effect is maximized. When it is bright or when there are many non-correlation components, the valley of the frequency characteristic has a shallow characteristic to prevent deterioration of resolution. In the case of the intermediate value, the characteristic becomes an intermediate value, and the noise reduction effect is moderate.

FIG. 3 (B) ... FIG. 3 (B) showing changes in the noise reduction effect for explaining the system of FIG.
Shown in. The vertical axis (NR amount) shows the noise reduction effect. The noise reduction effect decreases as the brightness increases and the decorrelation component increases, and the noise reduction effect increases as the contrast decreases and the amount of the decorrelation component decreases. Becomes In this way, the adaptive image filter of the present invention adapts to the input signal much more smoothly than in the past, and the image quality can be improved. FIG. 4 ... Another embodiment of the noise reduction circuit

This embodiment differs from the embodiment of FIG. 2 in that the limiter is also controlled. Therefore,
The same parts as those in the embodiment of FIG. 2 are designated by the same reference numerals. In addition to the above rules, rule… if… then rule 1… If there are many small uncorrelated components… increase the limiter rule 2… if there are few small uncorrelated components… decrease the limiter rule 3… if bright … Rule 4 to reduce the limiter… If it is dark… Add a rule to increase the limiter. The visual sensitivity is low in bright areas, so the limiter is small. Also,
Looking at the number (frequency) of small uncorrelated components is going to look at noise. That is, if the noise is small,
Noise is so small that it is not detected as a non-correlation component.If the noise is large, the noise is detected as a non-correlation component.Therefore, if the detection level is properly selected, it can be identified to some extent whether the image is noisy. it can. When there is a lot of noise,
The limiter level is increased to increase the range of noise amplitude that can be reduced. When the noise is small, the limiter level is reduced to prevent resolution deterioration. This operation is slow in time. Since the operation is performed by counting the number of small uncorrelated components, the operation is performed by integrating within a fixed time. Therefore, the response speed of adaptation is slow. In both cases of receiving a television broadcast and playing a VTR, a noisy source is naturally noisy while watching the program. That is, a slow response speed is sufficient and matches the control.

Here, the noise is to be detected by the signal itself, but a method of investigating in the sync portion (horizontal retrace line section or vertical retrace line section) is also conceivable. The detecting means 27 may detect the amount of the small non-correlation component in the blanking interval, hold the amount for the scanning time, and give it to the control signal generating means 29 as the amount of noise. Since there is no image signal in the blanking interval, noise discrimination is easier. FIG. 5 ... Still another embodiment of the noise reduction circuit

Another embodiment is shown in FIG. This is also an example of the luminance signal noise reduction circuit. This circuit is a system that detects the non-correlation component and controls the gain, a system that detects the correlation component and controls the gain, and a system that adds the outputs of both systems and subtracts the addition output from the input signal. Have a road. The system for detecting the non-correlation component has the same configuration as that of the embodiment shown in FIG. 2, and therefore the same functional parts are designated by the same reference numerals. The path for detecting the correlation component will be described. The input signal and the output of the 1H delay means 23 are added by the adding means 34. The output of the adding means 34 is limited by the limiter 35, and this limiter output is input to the adding means 37 via the variable gain amplifier 36. The adding means 37 adds the outputs of the variable gain amplifiers 30 and 36 and gives the result to the subtracting means 38. The subtraction means 38
Output terminal 2 by subtracting the input signal and the output of adding means 37
It outputs to 2.

That is, the sum (correlation component) and the difference (non-correlation component) of the signal delayed by the horizontal scanning time and the current signal are taken out, added through the limiter and the variable gain amplifier, respectively, and subtracted by the present means. It is the subtraction from the signal. Input signal and output of subtraction means 24, addition means 35
The respective outputs are input to the detection means 41, 42, 43, respectively, and feature extraction is performed. And each detecting means 41-43
Is output to the control signal generating means 44. The control signal generating means 44 determines the variable gain amplifier 3 by logical analysis.
The gain control signals for 0 and 36 are created and provided.

The image filter has a characteristic of gain = 1.0-{Ga (1.0-exp (-jωt)) + Gb (1.0 + exp (-jωt))} (2) in the range where the limiter is not applied. Here, Ga is the amplification factor of the variable gain amplifier 30 for the non-correlation component, and Gb is the amplification factor of the variable gain amplifier 36 for the correlation component. In this case, apply the rules shown in the following table.

Rule ... if ... Then Rule 1 ... If the amount of non-correlation component is large ... Reduce the amplification factor of Ga Rule 2 ... If the amount of non-correlation component is small ... Increase the amplification factor of Ga Rule 3 … If the non-correlation component occurs frequently… Reduce the amplification factor of Ga Rule 4… If the non-correlation component does not occur frequently… Increase the amplification factor of Ga Rule 5… If the amount of the correlation component is large… Gb Decrease the amplification factor Rule 6 ... If the amount of the correlation component is small ... Increase the amplification factor of Gb Rule 7 ... If the correlation component occurs frequently ... Reduce the amplification factor of Gb Rule 8 ... Do not occur the correlation component frequently Then ... Increase the amplification rate of Gb Rule 9 ... If it is bright (the brightness is high) ... Reduce the amplification rate of Ga and Gb Rule 10 ... If it is dark (the brightness is low) ... The amplification rate of Ga and Gb To make Ri has become complicated, since the control signal by performing a calculation of the fuzzy inference using the procedure given can be generated, computation itself is easy. In the case of the conventional image filter, the control signal generation procedure was designed using an adder, etc.
Although it is difficult to design when the complexity is so much, the method of the present invention has an advantage that the burden on the engineer is reduced because the result can be obtained by performing the calculation according to the fuzzy theory.
That is, the development period can be shortened, and as a result, an inexpensive system can be obtained.

The circuit of FIG. 5 is designed so that the noise reduction effect is greater than that of the circuit of FIG. In the case of the circuit of FIG. 4, only the noise included in the non-correlation component is reduced, but the method of the embodiment of FIG. 5 has an effect of reducing the noise of the correlation component as well. For signals with strong vertical correlation, such as vertical lines, the correlation component is large and the non-correlation component is small, but in the case of this type of image, Rule 5 reduces Gb and Rule 2
As a result, Ga becomes large and the correlation component becomes small and the non-correlation component becomes large in the signal, resulting in the same noise reduction effect as in the past. Furthermore, in an image with large vertical non-correlation such as diagonal lines, the correlation component is small and the hidden component is large, but in the case of this type of image, Rule 6 increases Gb and Rule 1 decreases Ga to reduce the signal It reduces the noise of uncorrelated components and does not suppress uncorrelated components with signals. This is a noise reduction capability that conventional noise reduction circuits do not have. Further, in a flat image, since both components are small, both Ga and Gb are increased by rules 2 and 6, and noise in the entire band is reduced. Furthermore, in random images, both Ga and Gb are reduced by rules 1 and 5 to suppress noise reduction and prevent resolution degradation. You can think about other rules in the same way. 6 ... Examples of frequency characteristics for explaining the circuit of FIG.

FIG. 6A shows a part of the frequency characteristic.
The frequency characteristics of the correlation path and the non-correlation path of Eq. (2) are comb-shaped as shown in the figure. When Ga = Gb = 0.5, the addition cancels the exp (-jωt) term and the sum of the extracted components becomes just 1.0, and the gain of Eq. (2) is 0.0
become. That is, when Ga = Gb = 0.5, the noise in the entire band is canceled. If Ga = 0.0 and Gb = 0.5, cancel the correlated noise. Ga = 0.5 Gb = 0.0 cancels uncorrelated noise. The present invention can take values between these values, and when Ga or Gb changes, not only the position of the crests and troughs of the frequency characteristic but also the height and the depth of the crests change. Thus, the noise reduction circuit of the present invention adapts to the input signal and flexibly changes to have the optimum filter characteristic. FIG. 6 (B) ... Waveform example for explaining the circuit of FIG.

In order to make the effect of the noise reduction circuit of FIG. 5 easy to understand, an example shown by a waveform is shown in FIG. 6 (B). Here, (a) to (d) are examples when the correlation is strong, and (e) to (d)
(h) is an example when the vertical decorrelation is strong. However, the effect of the limiter is disregarded for the sake of simplicity.

(A) An example of an input signal waveform of an image having a strong vertical correlation (b) An example of a waveform obtained by delaying an input signal of an image having a strong vertical correlation during a horizontal scanning period (c): A conventional noise reduction circuit (vertical Example of output waveform of (NR) (d) ... Example of output waveform of noise reduction circuit in Fig. 5 (e) ... Example of input signal waveform of image with strong vertical decorrelation (f) ... Input of image with strong vertical decorrelation Example of waveform obtained by delaying signal in horizontal scanning period (g) Example of output waveform of conventional noise reduction circuit (vertical NR) (h) Example of output waveform of noise reduction circuit in FIG.

First, (a) to (d) will be described. Expressed in the form of an image, this is a vertical line. Since the input signal and the delayed signal have the same waveform, the conventional noise reduction circuit outputs the waveform of (c) with no decorrelation component. Although (a) and (c) appear to have the same waveform, noise has no correlation, so (c) is a signal with reduced noise. The same applies to the case of the present invention, and the same waveform as in (c) appears as in (d).

Next, (e) to (h) will be described. When expressed in the form of an image, this becomes a diagonal line of about 45 degrees. The waveforms of the input signal and the signal delayed by the horizontal scanning time are
The shape is just inverted and looks like (e) and (f). In the case of the conventional noise reduction circuit, if the decorrelation cannot be detected as shown in (g), the signal is also removed as a price to reduce the noise, and if the decorrelation can be detected, , Because the noise reduction effect is zero or small,
A noisy signal like (e) will be output.
On the other hand, in the noise reduction circuit of FIG. 5, the correlation component is small and the non-correlation component is large, so that the rule 1 and the rule 6 are met, Ga is reduced, Gb is increased, and the correlation component has a valley. Change to a horseshoe filter. That is,
As shown in (h), a signal in which noise is skillfully reduced is output. As described above, it can be seen that this noise reduction circuit has an unprecedented noise reduction ability and is an extremely excellent adaptive noise reduction circuit. FIG. 7: Example of a cyclic noise reduction circuit

FIG. 7 also shows another embodiment applied to the luminance signal noise reduction circuit, and also shows a cyclic noise reduction circuit. The input signal from the input terminal 21 is input to the adding means 51 and added with the output of the variable gain amplifier 53. The output of the adding means 51 is input to the 1H delay means 52 and then to the variable gain amplifier 53. Therefore, the correlation component is output from the 1H delay means 52, and this component is gain-adjusted by the variable gain amplifier 5 and input to the subtraction means 55. Since the subtraction means 55 subtracts the output of the variable gain amplifier 54 from the input signal, a non-correlation component is obtained. The decorrelation component obtained from the subtraction amplifier 54 is the limiter 56 and the variable gain amplifier 57.
Is input to the subtraction means 58 via. Since the subtraction means 58 subtracts the output of the variable gain amplifier 57 from the input signal, the signal from which the non-correlation component (noise) is removed is finally derived to the output terminal 22. Here, the input signal and the decorrelation component from the subtraction means 55 are detected by the detection means 6 respectively.
1 and 62, the characteristics thereof are extracted and input to the control signal generating means 63. The control signal generating means 63 uses variable gain amplifiers 53, 54 and 57, and
Further, a control signal for the limiter 56 is created and supplied.

As for the gain of this circuit, assuming that the gain of the variable gain amplifier is g1, g2, and g3 from the left, the horizontal scanning time is t, the angular frequency is ω, and j = sqrt (-1). ,

[0051]

[Equation 1]

It becomes If the gain relation of the two variable gain amplifiers is always maintained in the relation of g2 = 1.0-g1, the gain of Eq. (4) becomes 1.0. That is, the gain of the noise reduction circuit for the correlation signal is 0 dB. (5) for uncorrelated signals. 10 ... Example of frequency characteristic for explaining the circuit of FIG.

If this frequency characteristic is shown in a graph, it has the form shown in FIG. If g1 and g2 are controlled while maintaining the relation of g2 = 1.0-g1, not only the depth of the valley of the comb but also the width of the valley changes. The wider the width, the wider the noise reduction capability. If g3 is also controlled, the width and depth can be controlled with considerable freedom.

In this figure, when g2 is increased, the width of the valley becomes wider, and making the width wider means that the noise reduction ability becomes larger, but the influence of the past line becomes more likely. .. When there is no vertical correlation (that is, when there is a change), the image is affected by the past and becomes a low-resolution blunt image. For uncorrelated signals, narrower troughs (ie smaller g1) are desirable. The circuit of FIG. 7 controls the width of this valley. If we call g2 the cyclic coefficient,

Rule ... if ... Then Rule 1 ... If the amount of decorrelation component is large ... Reduce the cyclic coefficient Rule 2 ... If the amount of decorrelation component is small ... Increase the circulation coefficient Rule 3 ... Decorate component Occurs frequently ... If the cyclic coefficient is reduced Rule 4 ... If the non-correlation component does not occur frequently ... Increase the cyclic coefficient Rule 5 ... If bright (high brightness) ... Reduce the cyclic coefficient Rule 6 ... Dark ( If the brightness is small) ... Apply a rule that increases the cyclic coefficient to obtain the desired image quality. FIG. 9 ... Another embodiment of the noise reduction circuit

FIG. 9 shows still another embodiment. This embodiment is basically the same as the embodiment of FIG. Therefore, the same functional parts as those in the embodiment of FIG. 4 are designated by the same reference numerals.
The difference is that a slicer 65 is provided between the subtracting means 24 and the variable gain amplifier 30 instead of the limiter.
The frequency characteristic is gain = 1.0 + exp (-jωt) (6) for a minute signal that does not reach the slice level of the slicer 65. The maximum value of gain is 2, but if the gain is halved so that it becomes 1, it is exactly the same as Eq. (1) when g = 0.5. FIG. 10 ... Example of three-dimensional noise reduction circuit

Another embodiment is shown in FIG. Although similar to the embodiment of FIG. 5, the delay means in FIG. 5 is a vertical noise reduction circuit for delaying the horizontal scanning time, whereas the delay means in this embodiment is a 1F delay means for performing a time delay of one frame. It is a three-dimensional noise reduction circuit using 66. Since they are basically the same, the same functional parts are designated by the same reference numerals. FIG. 11: Example of high frequency noise reduction circuit

FIG. 11 shows still another embodiment. This is also similar to the embodiment of FIG. 5, and in FIG.
The circuit portion formed by the subtraction means 24 is a band pass filter (BPF) 67 in this embodiment. If the pass band of the band pass filter 67 is designed to be a noise band to be removed, the noise in that band can be reduced. Other similar functional units are designated by the same reference numerals as those in FIG. gain = 1.0-BPF (f) (7)

Here, BPF (f) is the frequency characteristic of the bandpass filter. The same effect can be obtained even if the delay time of the delay means of FIG. 5 is set to several hundreds nsec instead of the band pass filter. FIG. 12: Configuration Example for Learning Neural Network Although the method of the present invention using fuzzy reasoning has been described above, the same effect can be obtained by using a neural network instead of fuzzy reasoning.

The neural network is known as a calculation system simulating human brain cells, and is constructed by arranging a large number of elements simulating cells in layers so that signals can be transmitted between the elements between layers. It is generally known that practical results can be obtained with about three layers. Depending on the coefficient of transmission between the elements,
That intelligence is remembered. When the coefficient of transmission between elements is set to an appropriate value, an appropriate output signal is output according to the stored knowledge with reference to the stored knowledge.

Although this technique is known, a method of applying this technique to an image filter is the present invention. As with fuzzy reasoning, neural network computations are enormous. The ratio of transmission between the elements is decided by the multiplier, and even a neural net which has nine terminals and can be said to be the minimum of a three-layer structure requires several hundreds of multipliers, and the multiplier requires a large area in the IC. Since it is a huge circuit required, it becomes a large-scale and expensive circuit. Also, there is a problem in terms of calculation speed.

FIG. 1 is also an embodiment thereof, and the partial neural network of the control signal generating means described so far is applied. When the extracted plurality of features are put into the neural network, the neural network generates an appropriate control signal. Since the feature extraction is performed, the resolution of quantization is small and the input data itself is small, so that the neural network can be configured and realized as an image filter.

Further, as described in the case of fuzzy inference, it is possible to adopt a table lookup system. The required memory is the same as in the case of fuzzy inference, and the amount can be easily incorporated in the IC. Neural nets are famous as learning devices that can be learned, but they are not suitable for practical use because they have a large learning function. If only the knowledge acquired by learning is used (that is, the coefficient is fixed), the table lookup method in which the learned knowledge is stored in the ROM can obtain the completely same result.

FIG. 12 shows an example of the structure of a learning process. A good control signal cannot be obtained simply by using a neural network, and a desired result cannot be obtained unless a properly trained neural network is used. In this example, the backpropagation method is devised so that it can be applied to the image filter of the present invention, so that learning can be correctly performed. In this figure, the part described as a test circuit is the image filter of the invention, that is,
For example, the circuit of FIG. This figure shows the learning method of the noise reduction circuit.

That is, the reference image and the reference noise are added by the addition means 80.
And input to the image filter to obtain an output image. At this time, the difference between the input image and the output image is obtained by the comparison means 81, and the control signal generation means (namely, neural network) inside the image filter is controlled so that the difference is minimized.
The control state is stored as a coefficient by the neural network. By doing this for several reference images and reference noise, and then repeating it, the neural network can be stored to generate control signals that reduce noise and do not reduce resolution.

The learned result may be stored in the memory and the neural network may be calculated as the coefficient of the multiplier, or may be stored in the memory as the relationship between the input and output of all states, and the control signal generating means may be looked up in a table. You can also turn it up.

Unlike the fuzzy method, the neural network method is difficult to change. If you change a part, you have to re-learn all of the neural network, but on the other hand, even in the case of processing that can not be ruled, it is optimized by giving a reference signal to the neural network and making it learn. There is an advantage that the output can be obtained. In the course described in FIGS. 1 to 11,
The noise was implicitly assumed to be random. This assumption is commonly used in general, but not always. For example, in the case of VTR reproduction, there is noise specific to the VTR. Some noise is concentrated in a specific band, such as noise caused by rotating system jitter and crosstalk noise of adjacent tracks. It is difficult to make clear rules for such applications. However, for example, in a VTR application, if no signal is recorded and reproduced, noise is sampled, and the neural network is learned to suppress the noise as the reference noise in FIG. 12, effective noise reduction is possible. The adaptive image filter using the neural network has such advantages. FIG. 13: Example of color signal noise reduction circuit

The embodiment of FIG. 13 is an embodiment applied to a noise reduction circuit for carrier color signals. This is also basically the same as in FIG. Therefore, the same functional parts as those in FIG. 5 are designated by the same reference numerals. However, in this embodiment, as the information for extracting the characteristics of the signal, the input terminal 91,
The detecting means 92 is provided, and the information of the luminance signal is also inputted to the control signal generating means 29. Since the NTSC system carrier color signal is selected as the frequency which is inverted every horizontal scanning time, the input signal and the signal delayed by the horizontal scanning time are added,
When it is small, subtraction means subtracts it from the input signal to reduce noise. Since the correlation component (same color) of the carrier color signal has the frequency that becomes the minimum when added in the time apart from the horizontal scanning period, the larger the output signal of the addition means, the larger the non-correlation component.

For example, using the following rules will result in a desirable color signal noise reduction circuit. In FIG. 13, the second
The input is a luminance signal. The color signal and the luminance signal often have a correlation, and the color signal often changes in the portion where the luminance signal changes. Rule 7 and rule 8 utilize this property. The luminance signal does not necessarily have to be used, but since the color signal is modulated, the use of the luminance signal facilitates extraction of features. The color change of rule 5 and rule 6 can be understood by examining the correlation of the times separated by 280 nsec.

Rule ... if ... Then Rule 1 ... If the amount of non-correlation component is large ... Reduce the amplification factor of the variable gain amplifier Rule 2 ... If the amount of non-correlation component is small ... Rule 3 ... If the non-correlation component frequently occurs ... Reduce the gain of the variable gain amplifier Rule 4 ... If the non-correlation component does not occur frequently ... Increase the amplification factor of the variable gain amplifier Rule 5 If the color change is severe ... Reduce the gain of the variable gain amplifier Rule 6 ... If the color change is severe ... Increase the gain of the variable gain amplifier Rule 7 ... If the luminance change is severe ... Reducing the amplification factor of the variable gain amplifier Rule 8 ... If the change in brightness is large ... Increasing the amplification factor of the variable gain amplifier Fig.14: Example of 3-line Y / C separation circuit

FIG. 14 shows an embodiment applied to a Y / C separation circuit. The composite video signal is introduced to the input terminal 101 and input to the delay unit 102 having a delay amount of one horizontal period. The output of the delay means 102 is further input to the delay means 103 having a delay amount of one horizontal period. The input signal and the outputs of the delay units 102 and 103 respectively include a band-pass filter (BP) that passes the components of the sub-carrier band.
F) is supplied to 104, 105, 106. The output of the bandpass filter 105 is inverted by the inverter 107. The adding means 108 adds the input signal and a signal obtained by delaying and inverting the input signal by one horizontal scanning time, and the adding means 109 adds the inverted signal and the output signal of the band pass filter 106. As a result, the line signals can be added using the three adjacent line signals to obtain the color signal. Signal of each line (output of BPF104,
The output of the inverter 107 and the output of the BPF 106) have their characteristic components extracted, and are input to the control signal generating means 116 to be operated to generate a control signal for controlling the variable gain amplifier 112. The output of the BPF 104, the output of the inverter 107, and the output of the BPF 106 are the intermediate value circuit 11
The intermediate value is extracted by inputting it to 0, and the output of the intermediate value circuit 110 is input to the intermediate value circuit 111. The intermediate value circuit 111 includes addition means 108, 10
The output from 9 is also input. This intermediate value circuit 111
The signal output from the variable gain amplifier 11 is a color signal.
It is entered in 2. The output color signal of the variable gain amplifier 112 is led to the output terminal 114 and also input to the adding means 113. The adding unit 113 obtains a luminance signal by subtracting the color signal from the composite video signal output from the delay unit 102, and outputs the luminance signal to the output terminal 115.

The luminance output extraction is added because the signal delayed by one horizontal scanning time is inverted after passing through the band pass filter (BPF). The control signal generating means 116 may be a fuzzy inference method or a neural network method. 15: Details of the control signal generating means of FIG.

The details of the control signal generating means are shown in FIG.
Inputs 1 to 3 are the output of the BPF 104 and the inverter 1, respectively.
It corresponds to the output of 07 and the output of the BPF 106. Input 1
3 is input to the delay means 131 to 133, and the outputs of the delay means 131 to 134 are further input to the delay means 134 to 136, respectively. Each delay means has a delay amount of, for example, half the period of the carrier color signal. The input 2 and the output of the delay unit 132 and the output of the delay unit 135 are input to the detection unit 317. The detection means 318 has an output of the delay means 131 and a delay means 132.
And the output of the delay unit 133 are input. The input 1 and the output of the delay unit 131, the output of the delay unit 132, and the output of the delay unit 136 are input to the detection unit 319. The detection means 140 has an output of the delay means 134,
The output and the input 3 of the delay means 132 are input. The output of each of the detection means 137-140 is the control signal generator 141.
Entered in.

In FIG. 15, since the data of three horizontally adjacent points are given to the detecting means 137, the horizontal correlation can be examined. Since the detection means 128 is connected to three vertically adjacent points, the vertical correlation can be examined. Since the detecting means 139 is connected to three diagonally adjacent points, it is possible to check the diagonal correlation. The detecting means 140 can check the diagonal correlation having an inverse gradient to that of the detecting means 139.

This Y / C separation circuit examines the correlation between adjacent scanning lines and performs Y / C separation by a comb filter between lines having a strong correlation. The so-called logical comb,
There is a comb filter called a dynamic comb. The conventional logical comb does not include a control signal generating means or a variable gain amplifier, but the present invention includes it.
Furthermore, we are trying to improve the separation performance by performing appropriate control by fuzzy reasoning or neural network.

It is known that the logical comb exhibits excellent separation characteristics in most images even in the conventional method, but there are some signals that it is not good at. An example thereof is shown in FIG. (a) to (e) are one set, and (f) to (j) are the other set.

(A) is an input signal, (b) is a signal before the horizontal scanning time, and (c) is a signal before the horizontal scanning time.
This waveform becomes a diagonal line in the image. Here, since the conventional logical comb separates only at three vertical points, the output waveform is as shown in (d). Since there is no correlation with each other, the first intermediate circuit becomes no signal and the output of the upper and lower adders is output in the second intermediate value circuit (that is, color signal), which is added to (b) and (d). A luminance signal such as is output. The luminance signal cannot be separated accurately and is lumped. The waveform obtained by inverting (d) is output as the color signal. In other words, it can be confirmed that the cross color appears.

The circuit of FIG. 15 has an adaptive control circuit to improve the separation performance. The detecting means and the control signal generating section of FIG. 15 control the variable gain amplifier 112 so as to prevent the rounding of the cross color and the luminance signal (resolution deterioration). In the case of the fuzzy inference control signal generating means, it can be realized by applying the following rules. If the output level of the detection means 130 in FIG. 15 is represented by D3 and the output level of the detection means 140 is represented by D4, the gain of the variable gain amplifier 112 is normally set to 1,

Rule ... if ... then Rule 1 ... If D3 is large and D4 is small ... Reduce the gain of the variable gain amplifier Rule 2 ... If D3 is small and D4 is large ... The gain of the variable gain amplifier is small By this, the diagonal line can be identified, and the variable gain amplifier is controlled according to the size. 16 ... Examples of waveforms for explaining the circuit of FIG.

An output example of FIG. 14 to which this rule is applied is shown in (e) of FIG. In the case of the waveforms (a), (b) and (c), the gain of the variable gain amplifier becomes small and the color signal is suppressed,
The luminance signal of (e) is output. Therefore, the cross color becomes small and the resolution degradation of the luminance signal is small.

When the amplitudes of the waveforms (a), (b) and (c) are small, rule 1 and rule 2 are only weakly applied.
Although it is not so different from the conventional circuit as compared with the circuit of FIG. 16, in such a case the originally output cross color is also small,
It doesn't matter.

Other examples are shown in (f) to (j). In the case of a conventional Y / C separation circuit that is not adaptive, the first intermediate value circuit
The waveform of (f) is taken out, and in the second intermediate value circuit (f)
Since the waveform obtained by halving is extracted, the luminance signal output is as shown in (i). That is, a false luminance signal is output where there is no luminance signal, and cross color appears even though there is no color. The output level of the detection means 137 in FIG.
If the output level of D1 and the detection means 138 is represented by D2, the gain of the variable gain amplifier is normally set to 1,

Rule ... if ... then Rule 3 ... If D1 is small and D2 is large ... Reduce the gain of the variable gain amplifier.

Then, the color output is suppressed, the cross color is reduced, and the luminance signal is reduced as the false signal as shown in (j). Among the shortcomings of the conventional logical comb, the biggest part is the two problems mentioned in the above example,

Rule ... if ... then Rule 1 ... If D3 is large and D4 is small ... Reduce the gain of the variable gain amplifier Rule 2 ... If D3 is small and D4 is large ... Rule 3 ... If D1 is small and D2 is large ... By controlling the amplification factor of the variable gain amplifier to be small, Fig. 16 becomes an excellent logical comb with very few defects. FIG. 17: Example of configuration for learning neural network by application to Y / C separation

FIG. 17 shows an example of a learning method when the control signal is composed of a neural network. It is basically the same as the case of FIG. A reference color signal and luminance signal are prepared, added by the adding means 150, and input to the test circuit. In this case, the test circuit is the Y / C separation circuit of the present invention, for example, the circuit of FIG. The comparator 151 compares the input luminance signal and the output luminance signal, and trains the neural network (control signal generator in the test circuit) so that the error output is minimized. Although an error output is obtained by comparing the luminance signals with each other in this figure, it is possible to learn by comparing the color signals with each other. FIG. 18 ... Example of 2-line Y / C separation circuit

FIG. 18 shows an embodiment applied to a 2-line comb filter. The degree of correlation between the two lines is checked, and the variable gain amplifiers 163, 1 are checked according to the degree of correlation.
65 is controlled. Upper stage variable gain amplifier 16
Reference numeral 5 controls the degree of band filter separation, and the variable gain amplifier 163 in the lower stage controls the degree of comb filter separation. The composite video signal of the input terminal 101 is input to the subtraction unit 162 via the 1H delay unit 161. The subtraction unit 161 performs a subtraction process between the input signal and the output of the 1H delay unit 161, and extracts a color signal component. The color signal component is input to the adding means 164 via the variable gain amplifier 163. The input signal is input to the adding means 164 under the gain control of the variable gain amplifier 165. The control signal generation means 166 includes an input signal and a 1H delay means 16
The output of 1 is input, and the correlation determination between two lines is performed. The low correlation means that the output of the subtraction unit 162 has many errors.

The output of the subtracting means 164 is input to the output terminal 168 as a color signal output via the bandpass filter 167, and is also input to the subtracting means 169. Subtraction means 1
69 derives a luminance signal to the output terminal 170 by subtracting the color signal from the input composite video signal.

According to this circuit, the first example of FIG. 16 cannot be improved, but the second example can be improved by this system. In the cases of (f) and (g) in Fig. 16, the comb filter separation outputs the waveform shown in (i) as the luminance signal, but the bandpass filter separation outputs the luminance signal (g).
Output a waveform like. Therefore, the circuit of the present invention
The waveform shown in (j) is output and the difference from (i) is improved. FIG. 19 ... Another embodiment of 3-line Y / C separation circuit

FIG. 19 also shows another embodiment of the 3-line Y / C separation circuit, which is an example in which no intermediate value circuit or correlator is used as compared with the embodiment of FIG. Therefore, blocks having the same functions as the blocks in FIG. 14 are designated by the same reference numerals.
Since there is no inverter on the output side of the BPF 105, the subtracting means 108a and 109a are used in this embodiment.
The outputs of the subtraction means 108a and 109a are subjected to gain control by the variable gain amplifiers 112a and 112b, respectively, and then added by the addition means 112c, and output from the output terminal 1 as a color signal output.
14 is derived. Further, this color signal is subtracted by the subtracting means 113.
The chrominance signal component input to a and included in the input composite video signal is removed. As a result, a luminance signal can be obtained at the output terminal 115.

This circuit examines the correlation between adjacent lines and controls the variable gain amplifiers 112a and 112b to use the comb filter having the stronger correlation. At this time, in the case of the pattern of FIG. 16, the point that a special rule is added and processed is similar to the case of FIG. Further, the control signal generating means of this embodiment may be configured as shown in FIG. 15 as in the case of the embodiment shown in FIG. FIG. 20: Embodiment to contour enhancement circuit

An example in which the present invention is applied to the contour emphasis circuit will be shown. In general, contour enhancement inevitably increases noise. This is because the noise and the signal cannot be distinguished from each other like the noise reduction circuit. Further, the contour emphasizing circuit also has an adverse effect that when a contour having a large amplitude is emphasized, an overshoot or a preshoot becomes too large, resulting in an image that is difficult to glare. For example, by applying the following rule, it is possible to realize an excellent contour enhancement circuit that produces little noise and does not give an unpleasant image.

Rule ... if ... Then Rule 1 ... If the uncorrelated component is very large ... Gain is made very small Rule 2 ... If the uncorrelated component is large ... Gain is made small Rule 3 ... Uncorrelated component is small If so ... Gain remains as it is Rule 4 ... If the non-correlation component is very small ... Reduce the gain Rule 5 ... If the image is bright ... Increase the gain Rule 6 ... If it is a dark image ... Reduce the gain As a result, it is possible to realize a contour emphasizing circuit in which noise is small and an overshoot is not excessive even with a large amplitude contour.

That is, in FIG. 20, the luminance signal supplied to the input terminal 180 is input to the secondary differentiating circuit 181, the contour component is extracted, the limiter 182 limits the input, and the variable gain amplifier 183 inputs it. It The output of the variable gain amplifier 183 is added to the original luminance signal in the subtracting means 184, and the contour-corrected signal can be obtained at the output terminal 185. The input signal and the output of the secondary differentiating circuit 181 are input to the feature detecting means 186 and 187, and the detection signal is input to the control signal generating means 188. The control signal obtained here is supplied to the control terminal of the variable gain amplifier 183.

Although only examples of controlling the variable gain amplifier (multiplier) have been described above, the control object of the present invention is not limited to the variable gain amplifier. For example, the contour emphasis circuit may change the center frequency to be emphasized. That is, if the center frequency of the secondary differentiating circuit in FIG. 20 is changed, the contour enhancing circuit changes the frequency characteristic. The frequency characteristic variable circuit can be realized by, for example, a transversal filter and changing the delay time of the filter. In the case of an analog circuit, it can be realized by changing the value of the variable impedance element. By examining the characteristics of the input signal, it is possible to easily know whether the signal components are concentrated in the high frequency band or the low frequency band. If a large number of bandpass filters are used, it is possible to finely divide the band and easily know which frequency component is large. If this is used, only the necessary band is adaptively emphasized. For example, the rules are as follows.

Rule ... if ... then Rule 1 ... If there are many low frequency components ... Shift the center of the emphasis frequency to the low range Rule 2 ... If there are many high frequency components ... Shift the center of the emphasis frequency to the high range To

[0097]

As described above, the adaptive image filter of the present invention can comprehensively judge the characteristics of the input image, optimize the circuit state, and output the signal in the optimum state. Since various characteristics are comprehensively judged, it has the advantage that detection errors and malfunctions are less likely to occur, and the optimization state is analog, much closer to the true optimum value than before. Can be in a state. Further, it is possible to adapt to a signal having no clear characteristic at an appropriate level without malfunction. This is an image filter that can significantly improve the image quality and can be applied to a higher level while significantly reducing the drawbacks of the conventional adaptive image filter.

In the case of the control signal generating means using the fuzzy inference, since the optimization program can be described by the rule, the operation by each extracted feature can be decomposed and rearranged and given, and when there are many features to be extracted. But it can be handled relatively easily. In addition, it is easy to deal with the case of changing the control, and it becomes possible to smoothly develop a higher-level adaptation.

In the case of the control signal generating means using the neural network, when changing the control, the neural network must be re-educated, and the control signal of the fuzzy reasoning is that the process is redone. Although it is inferior to the generation method, it has the advantage of being able to deal with cases where the shape of the rules cannot be clarified. Note that the present invention is not limited to the examples described above, but can be widely applied to image filters in general.

[Brief description of drawings]

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

FIG. 2 is a diagram showing an embodiment in which the present invention is applied to a vertical noise reduction circuit.

3 is a diagram showing an example of frequency characteristics for explaining the circuit operation of FIG. 2 and a diagram shown for explaining a change in noise reduction effect of the circuit of FIG.

FIG. 4 is a block diagram showing another embodiment in which the present invention is applied to a noise reduction circuit.

FIG. 5 is a block diagram showing still another embodiment in which the present invention is applied to a noise reduction circuit.

6 is a diagram showing an example of frequency characteristics for explaining the operation of the circuit of FIG. 4 and a diagram showing an example of waveforms for explaining an operation example of the circuit of FIG.

FIG. 7 is a diagram showing an embodiment in which the present invention is applied to a cyclic noise reduction circuit.

8 is a diagram showing an example of frequency characteristics for explaining the circuit operation of FIG.

FIG. 9 is a diagram showing still another embodiment in which the present invention is applied to a noise reduction circuit.

FIG. 10 is a diagram showing an embodiment in which the present invention is applied to a three-dimensional noise reduction circuit.

FIG. 11 is a diagram showing an embodiment in which the present invention is applied to a high frequency noise reduction circuit.

FIG. 12 is a diagram showing a configuration example for learning a neural network.

FIG. 13 is a diagram showing an embodiment in which the present invention is applied to a color signal noise reduction circuit.

FIG. 14 is a diagram showing an embodiment in which the present invention is applied to a 3-line Y / C separation circuit.

FIG. 15 is a diagram showing details of the control signal generating means of FIG.

16 is a diagram showing an example of waveforms for explaining the operation of the circuit of FIG.

FIG. 17 is a diagram showing a configuration example for learning a neural network when the present invention is applied to Y / C separation.

FIG. 18 is a diagram showing an embodiment in which the present invention is applied to a 2-line Y / C separation circuit.

FIG. 19 is a diagram showing still another embodiment in which the present invention is applied to a 3-line Y / C separation circuit.

FIG. 20 is a diagram showing an embodiment in which the present invention is applied to a contour emphasis circuit.

[Explanation of symbols]

13-15 ... Feature extraction means, 16, 29, 44, 63,
116, 166, 188 ... Control signal generating means, 17 ... Image filter, 18 ... Coefficient changing means, 23, 52, 161
... 1H delay means, 24, 26, 38, 55, 58, 16
2, 169, 184 ... Subtracting means, 25, 35, 56, 1
82 ... Limiters, 27, 28, 41-43, 61, 6
2, 92, 137 to 140, 186, 187 ... Detection means, 30, 36, 53, 54, 57, 112, 163,
165, 183 ... Variable gain amplifiers, 34, 37, 51,
80, 108, 109, 150, 164 ... Addition means, 6
5 ... Slicer, 66 ... 1F delay means, 67, 104-1
06, 167 ... Band pass filter (BPF), 81, 1
51 ... Comparison means, 102, 103, 131-136 ... Delay means, 107 ... Inverter, 110, 111 ... Intermediate value circuit, 141 ... Control signal generating section, 181, ... Secondary differentiation circuit.

Claims (11)

[Claims] 1. An image filter comprising an input terminal, an output terminal, and a coefficient varying means, wherein a plurality of feature extracting means for extracting features of an input signal and an output signal of the feature extracting means based on fuzzy inference are used. An image filter comprising: a control signal generating means for generating a control signal, wherein the coefficient varying means is controlled by a control signal from the control signal generating means. 2. An image filter comprising an input terminal, an output terminal, and a coefficient varying means, a plurality of feature extracting means for extracting features of an input signal, and an output of the feature extracting means constituted by using a neural network. An image filter, comprising: a control signal generating means for generating a control signal from a signal, wherein the coefficient varying means is controlled by the control signal from the control signal generating means. 3. The first or second aspect, wherein the control signal generating means is of a table lookup system.
The image filter described in the item. 4. A means for delaying the input signal by an integral multiple of a horizontal scanning time (including one frame), a means for synthesizing the input signal and the delayed signal to extract a non-correlation component, the non-correlation component 3. The image filter according to claim 1 or 2, further comprising means for varying the amount of 5. A delay unit for delaying the input signal by an integral multiple of a horizontal scanning time (including one frame), a unit for returning an output signal of the delay unit to an input side of the delay unit, and a delay unit for the delay unit. The image filter according to claim 1 or 2, further comprising: a unit that synthesizes an output signal and the input signal to extract a non-correlation component, and a unit that changes the amount of the feedback amount or the amount of the non-correlation component. .. 6. A filter comprising: a filter that allows a specific band of the input signal to pass; a varying unit that varies the amount of the signal extracted by the filter; and a unit that synthesizes the input signal and the output of the varying unit. The image filter according to claim 1 or 2, characterized in that. 7. A Y / C separation circuit for separating a color signal and a luminance signal from a composite video signal, wherein a variable gain means is provided in the color signal extraction system, and the variable gain means is used as a coefficient varying means. The image filter according to claim 1 or 2, wherein the image filter is configured to be configured as described above. 8. The input signal, a signal obtained by delaying the input signal by half the cycle of the color subcarrier, a signal obtained by delaying the input signal by the cycle of the color subcarrier, and horizontal scanning of these signals. Means for extracting nine types of signals, a signal before the time and a signal after the horizontal scanning time, and the information representing the characteristics of the input signal is extracted from the signals extracted by this means between the adjacent signals on the screen. The image filter according to claim 1 or 2, wherein the feature extracting means is configured. 9. A plurality of feature extracting means for extracting features of an input signal, a control signal generating means for generating a control signal from an output signal of the feature extracting means, and delaying the input signal by an integral multiple of a horizontal scanning time. Delay means, means for extracting a correlation component from the input signal and a delayed signal, means for extracting a non-correlation component from the input signal and a delayed signal, correlation component varying means for varying the amount of the correlation component, The apparatus further comprises a non-correlation component changing means for changing the amount of the non-correlation component, and a means for adding outputs of the correlation component extracting means and the non-correlation component extracting means, wherein the control signal generating means has the non-correlation component changing means. And an image filter configured to control the correlation component varying means. 10. An input terminal, an output terminal, a coefficient varying means,
And a plurality of feature extracting means for extracting features of the input signal, and a control signal generating means for generating a control signal from the output signal of the feature extracting means using a neural network and controlling the coefficient varying means with the control signal. An output of a reference noise source and a reference image signal source are combined and supplied to an image filter which controls the coefficient varying means by a control signal of a control signal generating means, and the output signal of the image filter and the reference are supplied. A learning method for an adaptive image filter, comprising: comparing with an output of an image signal source, and learning a coefficient of the neural network of the image filter so as to minimize the difference. 11. An input terminal, an output terminal, and a coefficient varying means,
And a plurality of feature extracting means for extracting features of the input signal, and a control signal generating means for generating a control signal from the output signal of the feature extracting means using a neural network and controlling the coefficient varying means with the control signal. The output of the reference noise source and the reference image signal source are combined and supplied to the Y / C separation circuit which is an image filter for controlling the coefficient varying means by the control signal of the control signal generating means. / C separation circuit and the output of the reference image signal source are compared, and the coefficient of the neural network of the image filter is learned so as to minimize the difference. Method of learning type image filter.