JP3486975B2 - Noise reduction apparatus and method - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a noise reduction device and a noise reduction method for reducing noise in an image signal containing noise.

[0002]

2. Description of the Related Art FIG. 15 shows an example of a conventional noise reduction device for reducing noise of an image signal containing noise by using various filter processing techniques such as three-dimensional processing technique. In the device shown in FIG. 15, an input signal such as a digital image signal is supplied to the terminal 211. This input signal includes a still image noise reduction filter 212 used when the input image in pixels is a still image portion, and a moving image noise reduction used when the input image is a moving image portion. It is supplied to the filter 213.
The filters 212 and 213 perform noise reduction processing according to the content of the input image, and the respective filter outputs subjected to the noise reduction processing are sent to the switching circuit 220.

The input image signal from the terminal 211 is further sent to the motion detection circuit 230, where the motion is detected in pixel units of the input image signal. This motion detection circuit 230
Outputs a motion detection signal, and by sending this motion detection signal to the switching circuit 220, the output from the filter 212 or the filter 213 is selected and controlled so as to be supplied to the output terminal 221 of FIG. ing. That is, when the motion detection circuit 230 determines that the input image signal is a still image, the output from the noise reduction filter 212 for a still image portion is supplied to the output terminal 221, and the motion detection circuit 230 moves the input image signal as a moving image. When it is determined that, the output from the moving image noise reduction filter 213 is supplied to the output terminal 221.

Therefore, the apparatus of FIG. 15 adaptively and appropriately reduces the noise component in response to the movement of the input image signal. The still image noise reduction filter 212 is an average value filter that smoothes in the time direction. This filter 21
2 is a frame memory 215 and coefficient multipliers 216 and 21.
7 and an adder 218. Frame memory 2
Reference numeral 15 outputs the previous frame data of the input image signal to the coefficient multiplier 216, where it is multiplied by the coefficient k (0 <k <1). The original frame data of the input image signal is multiplied by the coefficient (1-k) in the coefficient multiplier 217. The outputs from the coefficient multipliers 216 and 217 are supplied to the adder 218 and subjected to addition processing, and the addition result is filtered by the filter 212.
And is supplied to and stored in the frame memory 215.

The noise reduction filter 213 for the moving image section
Is an intermediate value filter such as a so-called median filter, which is a type of spatial filter. The intermediate value filter sets the level of a pixel having an intermediate level between two pixels including a certain pixel of interest to the level of the pixel of interest.

[0006]

As described above, the conventional image noise reduction apparatus employs a method of smoothing in the time direction, a method of using a filter such as a median in space, and the like. However, in any of the method of smoothing in the time direction and the method of using a filter such as a median in the space, the image is deformed by applying a filter to the image. It has the drawback of deterioration.

Therefore, for example, a filtering method has been proposed in which an edge of an image is determined and the edge portion is saved, but this is not optimal either.
In addition, the characteristics of noise generation often depend on the system that performs image processing, and the type of noise also differs. Therefore, it is difficult to reduce noise that matches the image shape and system.

In view of the above-mentioned circumstances, the present invention is a noise reduction apparatus and method in which deterioration of an image is small, and noise can be satisfactorily removed regardless of the shape, content and system of the image. Is intended to provide.

[0009]

The present invention has been proposed in order to solve the above-mentioned problems, and a noise reduction apparatus according to the present invention is a block forming means for dividing an input image signal into a plurality of blocks. , Class classification means for classifying each block into a plurality of classes by pattern classification using the blocked data, and outputting a class code corresponding to the classified class, and according to the class code of each block And a noise-reduced image signal output means for outputting a noise-reduced image signal corresponding to the block of the input image signal.

Here, the noise reduced image signal output means outputs the result of pre-learning with the data obtained by dividing the image containing noise into blocks and classifying the data and the data obtained by blocking the image with low noise into the class code. It is stored as a corresponding table. Further, the class classification means obtains a coefficient group of an expression representing the shape closest to the input image by using a high-order polynomial as a model expression, and performs class classification according to the coefficient group. Further, the class classification means performs class classification processing on the noise reduction filter output of the input image. Alternatively, the class classification means performs class classification processing on the difference between the input image and the noise reduction filter output thereof.

Further, the noise reduction method according to the present invention classifies each block into a plurality of classes by a blocking step of dividing an input image signal into a plurality of blocks and a pattern classification using the blocked data. A class classification step of outputting a class code corresponding to the classified class, and a noise reduced image signal for outputting a noise reduced image signal corresponding to the block of the input image signal according to the class code of each block And an output step.

In the low-noise image signal output step, the class is calculated from a plurality of learning values obtained by learning in advance with data in which an image containing noise is divided into blocks and classified into classes and data in which a low-noise image is divided into blocks. The learning value corresponding to the code is taken out. In the class classification step, a high-order polynomial is used as a model formula to find a coefficient group of an expression that represents the shape closest to the input image, and class classification is performed according to the coefficient group. Further, the class classification step classifies the noise reduction filter output of the input image. Alternatively, the class classification step performs class classification processing on the difference between the input image and its noise reduction filter output.

[0013]

According to the present invention, results obtained by learning in advance with data in which noise-containing images are divided into blocks and classified into classes and low-noise images into blocks (a table reflecting characteristics of image shapes) are obtained. ) Is saved. That is, since the learning table is learned between the input image and the real image of the low noise image signal, the characteristics of the system and the characteristics of noise are reflected. Therefore, the learning value corresponding to the class code according to the classification of the input image is extracted from this table (the noise-free image is fitted (mapped) instead of the noise-containing image), and the system is matched. Noise can be reduced. Also, because learning is performed using image shapes,
It is possible to perform noise reduction without deterioration of the image shape caused by filtering or the like. Further, once the table is created, the noise is removed only by the process of reading the table, so that the process is simplified.

[0014]

Embodiments of the present invention will be described below with reference to the drawings. As a first embodiment of the present invention, FIG. 1 shows a block diagram of a noise reduction apparatus for generating a noise-reduced image signal from an input image signal having a noise level higher than the noise-reduced image signal. There is.

The noise reducing apparatus of the first embodiment shown in FIG. 1 comprises a blocking circuit 2 for dividing an input image signal supplied through an input terminal 1 into a plurality of blocks, and the block data. To a plurality of classes and outputs a class code corresponding to the class, which is a class code generation circuit 3 and a low noise image signal corresponding to the input image signal according to the class code. The ROM table 4 is a noise reduction image output means.

A specific example of blocking by the blocking circuit 2 will be described with reference to FIG. As shown in FIG. 2A, the input image signal is shown in the form of a two-dimensional array in which the pixels 22 are arrayed in the row and column directions.
A three-dimensional image is formed. In the example shown in FIG. 2A, the blocking circuit 2 has three pixels in the upper left corner centered on the pixel 22a.
The first block is generated in the form of a block 24 of × 3 pixels. The blocking circuit 24 also generates the next block signal 26 centered on the pixel 22b as shown in C of FIG. Essentially, the blocking circuit 2 divides the input image signal into block signals of a predetermined size. This block is centered on each pixel in the input image signal except pixels along the edges of the image. Here, the blocks of A and B in FIG. 2 are arranged as pixels in 3 rows and 3 columns, but other arrangements may be used instead. For example, in one exemplary embodiment of the invention, the input image signal may be divided into blocks of 8x8 pixels, while in other examples the blocks may have different sizes.

A block signal, which is a blocked signal, is supplied to the class code generation circuit 3 by the blocking circuit 2, and the class code generation circuit 3 generates a class code according to each supplied block signal. Each class code sequentially identifies low noise pixel data corresponding to the central pixel (22a, 22b, etc.) of each corresponding block signal in which the code is generated. The low noise pixel data is obtained from the low noise image signal according to the training process described later.

The ROM table 4 stores low noise pixel data as a table of values associated with each class code. The class code is used as a read address for outputting the corresponding low noise pixel data from the ROM 4, for example. Alternatively, the read address is generated based on the class code. Therefore, for each block signal generated by the blocking circuit 2 and supplied to the class code generating circuit 3, low-noise pixel data is output from the ROM 4 to the output terminal 5, and as a result, an image signal with reduced noise Is generated. The block signal is generated for each pixel except the edge of the image of the input image signal. For the pixels at the edge,
When the image signal is displayed, it usually does not appear and therefore does not need to be processed.

The ROM table 4 stores a result (learning value) obtained by performing learning in advance on the data obtained by dividing the image signal containing noise into blocks and classifying the data and the data obtained by dividing the image signal with low noise into blocks. ing. Further, the class code generation circuit 3 uses a high-order polynomial, which will be described later, as a model formula to obtain a coefficient group of an expression that represents the shape closest to the input image, and classifies the class according to the coefficient group.

Here, learning of a low noise image and generation of a ROM table will be described. First, an example of a learning method for the embodiment shown in FIG. 1 will be described with reference to FIGS. 2A to 2D and the flowchart of FIG. Learning is performed by using a relatively noisy input image signal (step S1 in FIG. 3) and by using a low noise image signal (step S2).

In order to learn a noise-free image, it is necessary to first generate the noise-free image. FIG. 4 shows a still image without noise or a still image for n frames for generating a low noise image signal. That is, in this embodiment, the image signal of the low-noise still image is obtained by averaging still images of n frames in the time direction t as shown in FIG.

In step S3, the input image signal is divided into blocks. This blocking step can be realized by various methods. For example, it is a method of simply extracting the pixels of a specific block related as a block signal. In step S4, the pixels in the low noise image signal are separated at the position corresponding to the block formation in step S3. 2A and 2B, in relation to the block 24 of the input image signal, the center pixel 25 of the corresponding block 25a of the low noise image signal is separated. 2C,
Referring to D, in relation to the block 26 of the input image signal, the center pixel 27 of the corresponding block 27a of the low noise image signal is isolated. This process is repeated for all blocks and pixels.

The input image signal divided into blocks in step S3 is sent to the next step S5, and in this step S5, clustering is performed to classify the block data of the input image into classes. Note that class classification or clustering is pattern classification using blocked image data. This step S
In 5, class classification can be performed using adaptive dynamic range coding (ADRC) as described below.
When this ADRC is performed, the pixel data is divided into blocks as described above, a dynamic range based on the maximum pixel level and the minimum pixel level is defined for each block, and the dynamic range is encoded to encode each pixel. The number of quantization bits for each block based on range is selected. The minimum pixel level is subtracted from the level of each pixel in the block, and the rest is encoded by the selected number of quantization bits.

In step S6, which follows the processing of steps S5 and S4, training (learning) is performed between the blocks of the input image and the low noise image signal. In the learning step in step S6, each pixel 25, 27 of the noise-reduced image signal obtained in step S4,
The class code of the corresponding block signal of the input image signal is associated. Specifically, the pixels of the low-noise image signal corresponding to the same class code are averaged to obtain the noise-reduced pixel data as the optimum value or the representative value corresponding to the class code.

In step S7, it is determined whether or not learning has been repeatedly performed on all the blocks forming the image to the extent that all classes are sufficiently learned.
If it is determined NO in step S7, the process returns to step S3 for the input image and step S4 for the low noise image signal. If YES is determined in step S7, the process proceeds to step S8.

In step S8, a mapping table is generated with the optimum value learned for each class. This is ROM
It becomes a table. There are some patterns depending on how to learn the block of the low noise image signal corresponding to the block of the input image. For example, there may be a case where learning is performed by a block of an input image and a block of a low noise image signal and output, or a case where learning is performed by a pixel and output. Further, there are a case where the input image itself is learned as data to be learned, and a case where a difference amount from the filter is learned.

Further, although actual data can be used as the class in this embodiment, it is possible to use the data compression result or the like for class classification in consideration of the memory capacity considering the feasibility. For data compression, for example, adaptive dynamic range coding (ADRC) or differential quantization (DPC) according to a dynamic range of an image signal described later is used.
M), vector quantization (VQ), etc. are considered.

The ADRC means that the digital image signal is divided into blocks, the dynamic range defined by the maximum value and the minimum value of a plurality of pixel data included in this block is calculated, and the dynamic range is calculated in block units. Each pixel data in the block is encoded by the number of allocated bits adapted to the range. More specifically, the ADRC is a dynamic range (difference between maximum level and minimum level within a block) defined by maximum and minimum values of a plurality of pixels included in a predetermined block within one frame (or field). ) Is determined for each block, and the variable quantization bit number adapted to this dynamic range is determined, and the input data after the minimum value removal is encoded (requantized, that is, compressed quantum The dynamic range is evenly divided according to the number of encoded bits, and each pixel in the block is encoded to the code of the closest level) to obtain an encoded code, and the dynamic range information and the maximum and minimum values Is used as the additional code, and the encoded code and the additional code are obtained.

Furthermore, the method of realizing clustering includes
The following method is used. Here, the simplest method for class division is, for example, a method in which a bit sequence of learning data in a block is directly used as a class number. However, this method requires a huge capacity of ROM. Therefore, in the present embodiment, the ADRC is used for class division according to the signal pattern, and the number of classes is effectively reduced while preserving the characteristics of the signal pattern.

Here, for example, the learning signal pixels a, b, c
Taking the prediction to the original signal pixel A based on the above as an example, let the levels of the learning signal pixels a, b, and c be x ₁ , x ₂ , and x ₃ , respectively.
Also, p-bit ADRC is applied to the data of x ₁ , x ₂ and x _3.
Each re-quantized data of the result of q _1, q _2, q ₃
And the dynamic range is DR and the minimum value is min.
And

At this time, the class number (class code) class of this block is defined by the equation (1).

[0032]

[Equation 1] Originally, ADRC was a VTR (video tape recorder).
This is an adaptive requantization method developed for high-efficiency coding for video signals. Since a local pattern of signal level can be efficiently expressed by a short word length, this embodiment classifies it as a signal pattern class classification. It is used to generate the code.

Returning to FIG. 1, a configuration example of the noise reducing apparatus of the first embodiment for reading out data from the ROM table obtained by learning as described above and removing noise will be described. The noise reduction apparatus of FIG. 1 shows a configuration in the case of outputting pixels from which noise has been removed from an input image. In FIG. 1, the digital image signal input to the input terminal 1 is sent to the blocking circuit 2. The blocking circuit 2 divides the supplied digital image signal into a plurality of blocks composed of a plurality of pixel data.

The block data from the blocking circuit 2 is sent to the class code generating circuit 3. The class code generation circuit 3 obtains a class by referring to the supplied two-dimensional level distribution pattern of the block and generates the class code. This class code is sent to the ROM table 4 as address data. The RO
The learning value (that is, low-noise pixel data) as described above is stored in the M table 4, and the learning value is output corresponding to the class code (address). The learning value read from the ROM table 4, that is, the low-noise pixel data is output from the output terminal 5.

Next, as a second embodiment of the present invention, FIG.
Shows the configuration in the case where the learning value in block units is stored in the ROM table 9. 5, the configuration from input terminal 6 to class code generation circuit 8 is the same as that from input terminal 1 to class code generation circuit 3 in FIG. Here, the ROM table 4 in FIG. 1 outputs data that is not in block units, whereas the ROM table 9 in FIG. 5 outputs data in block units.

The low noise block unit data output from the ROM table 9 is stored in the block decomposing circuit 10.
Sent to. That is, since the output of the ROM table 9 is in units of blocks, the block decomposition circuit 10 performs block decomposition, and the decomposed data becomes output data of the low noise image signal from the output terminal 11. In the second embodiment, the ROM table 9 and the block decomposition circuit 10 constitute a low noise image signal output means.

Here, the learning for generating the table stored in the ROM table 9 is performed by the same method as that described for the ROM 4 of FIG. 1 with FIG.
However, the ROM 9 is different from the ROM 4 in that all data in block units are stored in the ROM 9 in correspondence with each class code. Regarding the learning method of the apparatus of FIG. 5, FIG.
Will be described with reference to.

6A shows a two-dimensional pixel array of the input image signal containing noise, and FIG. 6B shows the same portion as the input image signal of FIG. 6A and shows a low noise level. 2 illustrates a two-dimensional pixel array of a low-noise image signal that it has. As shown in A of FIG. 6, the two-dimensional pixel array is divided into block signals 32 of a predetermined size. Although a block signal of 3 × 3 pixels is shown in FIG. 6 as an example, a block signal of 8 × 8 pixels and block signals of other sizes and shapes can be used. The low noise image signal of FIG. 6B has a plurality of pixels 36 divided into respective block signals 38 corresponding to the respective block signals 32 of the input image signal of FIG. 6A. Therefore, FIG.
, Each block signal 32 of the input image signal is used to generate a class code, and the corresponding block signal 38 of the low noise image signal is related to the class code of the corresponding block 32. Will be remembered. Similar to the case of the learning procedure of FIG. 3, the block signals 38 having the same class code are averaged to obtain the noise reduction block signal as the optimum value or representative value corresponding to the class code. The block signal as the optimum value thus obtained is stored in the ROM 9 with the corresponding class code as an address.

Next, FIG. 7 shows, as a third embodiment of the present invention, the output of a commonly used filter for noise removal or noise reduction and the output of the blocking circuit 13 similar to FIGS. An example of the case where learning is performed first with respect to the difference of is shown. That is, the class code generation circuit 1 of the noise reduction apparatus of the third embodiment shown in FIG.
6, the input image blocked through the blocking circuit 13 and the block image are processed by the noise reduction filter 1
The class classification process is applied to the difference from the output filtered through 4.

In FIG. 7, a digital image signal is supplied to the input terminal 12 similarly to the input terminals 1 and 6 of FIGS. 1 and 5, and is blocked by the blocking circuit 13 similar to those of FIGS. Is done. The blocking circuit 13
The output of is sent to the same noise reduction filter 14 that is normally used. The output filtered by the filter 14 is sent to the adder 15 as a subtraction signal. At this time, the output from the blocking circuit 13 is also supplied to the adder 15 as an addition signal, so that the adder 15 can obtain the difference between the filtered block data and the unfiltered block data. become.

The output of the adder 15 is sent to the class code generation circuit 16. The class code generation circuit 16
However, the class is obtained by referring to the pattern of the two-dimensional level distribution of the output block of the adder 15 and the class code is generated. This class code is sent to the ROM table 17 as address data. The ROM table 17 stores learning values learned by difference data between low-noise pixel data and filtered data of the low-noise pixel data. Therefore, the class code (address) is stored in the class code (address). The learning value will be output correspondingly.

The learning value output from the ROM table 17 is also sent as an addition signal to the adder 18 to which the output of the filter 14 is supplied as an addition signal. As a result, low-noise pixel data is output from the adder 18 and output from the output terminal 19. Therefore, in the device of the third embodiment, the RO
M table 17 and adder 1 to which filter output is supplied
And 8 constitutes a low noise image signal output means.

Here, a method for learning the apparatus of FIG. 7 will be described with reference to the flowchart of FIG.
In the learning procedure shown in FIG. 8, the input image signal is divided into blocks in step S10, and the low noise image signal is divided into blocks in step S11. These steps of blocking are the same as step S3 shown in FIG.

In step S12 after step S10 and step S13 after step S11, the noise removal filter or noise reduction filter 14 shown in FIG. 7 is applied to the input image signal and the low noise image signal. The same filtering as that of is applied. The image signal whose noise is reduced by filtering in these steps S12 and S13 is
In S14 and S15, they are subtracted from the image signal before the filtering process. Step S1 following step S14
In step 6, the clustering for classifying the block data of the image signal into classes is performed on the differential image signal generated in step S14 as in step S5 of FIG. In step S20, which proceeds after the processing of step S16 and step S15, the class code generated in step S16 and the difference image data obtained from the block signal of the low noise image signal are associated and learning is performed.

In the next step S22, it is determined whether or not learning has been repeatedly performed on all the blocks forming the image to the extent that all class learning is sufficiently performed. If it is determined no in step S22, the process returns to step S10 for the input image and to step S11 for the low noise image. If YES is determined in step S22, the process proceeds to step S24, and a mapping table is generated with the optimum value learned for each class. This becomes the ROM table 17 in FIG.

In the noise reducing apparatus of the third embodiment shown in FIG. 7, although the difference is classified, the noise reducing filter 14 for the input image is used.
It is also possible to classify the output of. Next, a fourth embodiment of the present invention will be described with reference to FIG. In the fourth embodiment shown in FIG. 9,
After the input image signal supplied to the terminal 30 is blocked by the blocking circuit 32, it is sent to a general noise reduction or noise removal filter 34 to perform filtering for reducing or suppressing the noise level. Each block signal obtained by filtering by the filter 34 is supplied to the class code generation circuit 38, and a class code for identifying each block signal is generated. This class code is ROM as address data
It is sent to the table 40, the corresponding low noise pixel data is read out, and is taken out from the output terminal 42.

The learning procedure in the case of the fourth embodiment of FIG. 9 will be described with reference to FIG. In step S30 of FIG. 10, the input image signal is divided into blocks, and in step S32, the low noise image signal as described above is divided into blocks. These steps of blocking are the same as step S3 shown in FIG.

In step S34 after step S30, the same filtering as that performed by the noise reduction filter or noise removal filter 34 of FIG. 9 is applied to the input image signal and the low noise image signal. It The image signal whose noise has been reduced by filtering in these steps S34 is processed in the next step S34.
In step S36, the filtered low noise image signal obtained in step S34
Similar to 5, clustering is performed to classify the block data of the image signal into classes. In step S40 that follows the processing of steps S36 and S32,
The class code generated at 36 and the corresponding block signal of the low noise image signal generated at step S32 are matched or associated to generate a learning data set. That is, when the pixel data is stored in the ROM, the class code is associated with the pixel of the corresponding block signal of the low noise image signal, and when the block signal is stored in the ROM, the class code and the low noise image signal are stored. Correspondence with the corresponding block signal is made.
Pixels and blocks with the same class code are processed in step S40.
By averaging in, the noise-reduced pixel or block data as the optimum value or representative value corresponding to the class code is obtained.

In the next step S42, it is determined whether or not learning has been repeatedly performed on all the blocks forming the image to the extent that all classes are sufficiently learned. When it is determined NO in step S42, the process returns to step S30 for the input image and step S32 for the low noise image. If YES is determined in step S42, the process proceeds to step S44 to generate a mapping table with the optimum value learned for each class. This becomes the ROM table 40 in FIG.

Next, FIG. 11 shows, as a fifth embodiment of the present invention, a noise reducing apparatus in which a noise-reduced image signal is generated using a set of coefficient data read from the ROM 56. ing. In the noise reduction device shown in FIG. 11, the input image signal having noise supplied to the input terminal 50 is divided into blocks by the blocking circuit 52, and the block code generation circuit 54 and the noise reduced image data generation circuit 58 respectively. Sent. The class code generation circuit 54 generates a corresponding class code according to each block signal from the blocking circuit 52, and supplies this class code as an address to the ROM 56.
This class code identifies the low noise level image data of the corresponding part of the image.

Here, the coefficient data is stored in the ROM 56 as described above. That is, the ROM 56
Stores a plurality of coefficient sets at each address corresponding to the corresponding class code. The coefficient data set has a plurality of multiplication coefficients, and each multiplication coefficient is multiplied by the corresponding value of the block signal from the blocking circuit 52 in the noise reduction image data generation circuit 58 to obtain low noise image data. Used for. The low noise image data from the noise reduced image data generation circuit 58 is output to the output terminal 60.
In the form of noise-reducing pixels.

Here, the learning of the apparatus shown in FIG. 11 can be performed in the same manner as the method shown in FIG. 3 for learning the apparatus shown in FIG. However, the difference is that learning is performed to generate the above-mentioned coefficient data. Next, FIG.
Is a block circuit 66 according to a sixth embodiment of the present invention.
The noise reducing device is shown in which the block signal from the noise reduction block is sent to the class generation circuit 68 and the noise reduction block generation circuit 72, and the output signal from the noise reduction block generation circuit 72 is transmitted to the block decomposition circuit 74.

In the noise reduction device shown in FIG. 12, the input image signal containing noise supplied to the input terminal 64 is divided into blocks by the blocking circuit 66, and the class code generation circuit 68 and the noise reduction block generation circuit are provided. 72 respectively. Class code generation circuit 68
Then, a corresponding class code is generated according to each block signal from the blocking circuit 66, and this class code is supplied to the ROM 70 as an address. This class code identifies the low noise level image data of the corresponding part of the image.

Low noise image data is stored in the ROM 70.
It is stored in the form of coefficient data sets associated with block signals. The ROM 70 reads each coefficient data set according to each class code from the class code generation circuit 68, and the coefficient data set is sent to the noise reduction block generation circuit 72. Corresponding block signals from the blocking circuit 66 are supplied to the noise reduction block generation circuit 72, and each noise reduction block signal is generated.

The noise reduction block generation circuit 72 supplies each noise reduction block signal to the block decomposition circuit 74, and the block decomposition circuit 74 receives the input data from the field of the noise reduction pixel signal composed of a series of pixel lines. Alternatively, it is rearranged to generate a frame and supplied to the output terminal 76. Here, the learning of the apparatus of FIG.
This can be performed in the same manner as the method shown in FIG. 3 for learning the apparatus of FIG. 1 described above. However, similar to the embodiment of FIG. 11 described above, the difference is that learning is performed to generate the coefficient data described above.

Next, an example of learning the apparatus shown in FIG. 12 will be described. Here, it is assumed that each pixel from the noise reduction block generation circuit has a value y _0s . This pixel value y _0s has coefficients w _{1s to} w
_ns and can be generated by the combination of the pixel values x _IN1 ~x _INn of block signal from the blocking circuit _{66, y 0s = w 1s x} IN1 + w 2s x IN2 + ··· + w ns x INn (2) of such Be related to. Here, s is an index for identifying the arbitrary pixel value y _0s .

Learning is performed on a plurality of signal data for each class. To generate the coefficient data for storage in ROM 56, each corresponding pixel of the low noise block signal corresponding to the same class code is used with all pixel values of the corresponding block signal of the input image signal. The method for obtaining the coefficient data is derived as follows.

It can be estimated that each corresponding pixel of the low noise image signal is equal to a value represented by a polynomial obtained by multiplying each coefficient value w _{1s to} w _ns by the corresponding pixel value of the block signal of the input image signal. . That is expressed by the equation of _{y k = w 1 x k1 +} w 2 x k2 + ··· + w n x kn (3). Here, n is the number of pixels of each block signal of the input image signal, y _k represents the corresponding pixel value of the corresponding block signal k of the low noise image signal,
x _{k1 to} x _kn represent the pixel data values of the corresponding block signals of the input image signal, and w _{1 to} w _n are coefficients to be obtained by learning. Now, let the number of block signals be m
In the case of m> n, w ₁ to w _n are not uniquely determined, and therefore the element of the error vector e is e _k = y _k − {w ₁ x _k1 + w ₂ x _k2 + ... + wnx _kn } (k = 1,2, ..., m) (4) is defined and the coefficient that minimizes the equation (5) is obtained. This is the so-called least squares method.

[0059]

[Equation 2] Here, the partial differential coefficient by w _i of equation (5) is obtained as in equation (6).

[0060]

[Equation 3] Since each w _i may be determined so that the above equation (6) is set to 0, the following equations (7) and (8) are obtained.

[0061]

[Equation 4]

[0062]

[Equation 5] (9) is obtained by using a matrix as

[0063]

[Equation 6] This equation (9) can obtain the prediction coefficient as the optimum value by solving for w _i using a general matrix solution method such as the sweeping method. In this embodiment, this prediction coefficient is stored as data (ROM table) of the address class of the ROM 70.

By the way, the learning method of the apparatus shown in FIG. 11 is similar, but the difference is that only one set of coefficient data values is obtained and stored according to each class code. This is because only the low noise pixel is generated by the noise reduced image data generation circuit 58 according to each block signal generated by the blocking circuit 52. Next, FIG. 13 shows, as a seventh embodiment of the present invention, a class code is generated for the difference between the output of the noise removal or noise reduction filter that is normally used and the output from the blocking circuit 82. An example is shown in which the noise-reduced image signal is obtained by using the data obtained by reading the ROM 92 using the class code as an address.

That is, in the class code generating circuit 90 of the noise reducing apparatus of the seventh embodiment shown in FIG. 13, the input image blocked by the blocking circuit 82 and the noise reduction filter 84 are applied to the block image. The output obtained by sending the filtered output through the subtractor 86 to the subtractor 86 is subjected to class classification processing.

In FIG. 13, the input image signal supplied to the input terminal 80 is sent to the blocking circuit 82 to be blocked. The output from the blocking circuit 82 is sent to a noise reduction filter 84 and a subtractor 86, which are the same as those normally used. The output filtered by the filter 84 is sent to the subtractor 86 and is subtracted from the output of the blocking circuit 82 to obtain the difference between the filtered block data and the unfiltered block data.

The output from the subtractor 86 is sent to the class code generation circuit 90. The class code generation circuit 90
However, the class is obtained by referring to the two-dimensional level distribution pattern of the output block of the subtractor 86, and the class code is generated. This class code is sent to the ROM table 92 as address data. The ROM table 92 stores the learning value obtained by learning with the difference data between the low-noise pixel data and the filtered data of the low-noise pixel data. Therefore, the class code (address) is stored in the learning code. The learning value will be output correspondingly.

The ROM table 92 stores a plurality of coefficient data sets at respective addresses corresponding to the corresponding class codes. Each coefficient data set has several multiplication factors, and these multiplication factors are
The code is for multiplying a corresponding value included in the generated difference block signal to generate a value of the corresponding low noise difference block signal.

The coefficient data set read out from the ROM 92 according to each class code is the difference data generation circuit 9
4 and the difference data generation circuit 94 is supplied with the difference block signal generated by the subtractor 86. The difference data generation circuit 94 is a ROM
The difference data of each low noise image data is generated by multiplying each coefficient of the coefficient data set from 92 and the corresponding value of the difference block signal from the subtractor 86 and adding. The low-noise sub-data generated by the differential data generation circuit 94 is sent to the adder 98, and the adder 98 adds the low-noise sub-data to the output signal from the filter 84 to generate pixels of the noise-reduced image signal. The noise-reduced image signal from the adder 98 is output to the output terminal 10.
It is taken out from 0.

The learning of the apparatus of FIG. 13 can be performed by the same method as that shown in FIG. 8 in the case of the apparatus of FIG. 7, except that the data from the differential block signal is used instead of step S20 of the learning process. The difference is that the previously generated coefficient data is generated. Next, FIG. 14 shows a noise reducing apparatus as an eighth embodiment of the present invention. In FIG. 14, the input image signal is supplied to the input terminal 104, and this is sent to the blocking circuit 106. The blocking circuit 106 divides the input image signal into a plurality of block signals, and each block signal is supplied to a noise removal filter 108 suitable for reducing noise in the image signal. The noise removal filter 108 generates a filtered block signal and sends each signal to the class code generation circuit 112.

The class code generation circuit 112 generates a class code corresponding to each of the above-mentioned filtered block signals. Each class code identifies low noise image data that has a lower noise level than the block signal from which the corresponding filtered block signal was formed. The class code generation circuit 112 stores each code in the ROM 1
Supply to 14. This ROM 114 stores each of a plurality of coefficient data sets at each address corresponding to the corresponding class code. The ROM 114 reads the coefficient data set according to each class code and sends it to the noise reduced image data generation circuit 118. The noise reduced image data generation circuit 118 is supplied with the filtered block signal from the noise removal filter 108 for generating the class code. The noise-reduced image data generation circuit 118 multiplies each coefficient of the coefficient data set by the corresponding value of the filtered block signal and adds them,
Generating low noise pixels of the noise reduced image signal. The noise reduced image data generation circuit 118 has an output terminal 120.
To sequentially supply pixels.

As described above, in the noise reduction of the embodiment of the present invention, the image classification including the blocked noise is performed, and the noise reduction is performed from the table previously learned according to the class (that is, the signal pattern). It is possible to obtain a noise-free image by outputting the data that acts on the image and generating a block from which the noise is removed. Therefore, there is little deterioration of the image, and the image shape, contents and system are not affected. Instead, the noise can be removed well.

That is, in the digital image signal,
When trying to remove noise from an image in which noise is mixed, the image is usually deformed by filtering, but in the embodiment of the present invention, instead of using a filter,
Pattern (class) using blocked image data
The noise is removed by performing classification and extracting pre-learned noise-free data from the table according to this classification.

As a result, according to the noise reduction device of this embodiment, it is possible to obtain a noise-reduced image for an image picked up by, for example, a television camera. The present invention is not limited to the above-described embodiments, and various memory devices can be used instead of the ROM, for example. Further, the various embodiments described above can be realized as a hardware circuit circuit such as an ASIC, using a programmable device such as a microprocessor, a microcomputer, a digital signal processor, or an apparatus in which the hardware circuit and the programmable device are combined. .

Although specific embodiments of the present invention have been described in detail with reference to the drawings, the present invention is not limited to these strict embodiments and does not depart from the gist of the present invention. It goes without saying that various modifications and changes can be made within the range.

[0076]

As described above, in the present invention, the result of learning in advance is stored with the data obtained by classifying the image containing noise into the blocks and classifying the data and the data obtained by blocking the image with low noise into the input. Divide the image into multiple blocks, classify the blocked data into multiple classes, obtain a class code corresponding to that class, and output a low-noise image corresponding to the input image according to this class code. Therefore, the deterioration of the image is small, and the noise can be satisfactorily removed regardless of the shape, content, and system of the image.

[Brief description of drawings]

FIG. 1 is a block circuit diagram showing a schematic configuration of a noise reduction device according to a first exemplary embodiment of the present invention.

FIG. 2 is a diagram used for explaining a configuration of a block signal in the first embodiment of the present invention.

FIG. 3 is a flowchart for explaining learning in the first embodiment of the present invention.

FIG. 4 is a diagram for explaining prediction of an original signal image from a learning signal image.

FIG. 5 is a block circuit diagram showing a schematic configuration of a noise reduction device of a second exemplary embodiment of the present invention.

FIG. 6 is a diagram used for explaining a configuration of a block signal in the second embodiment of the present invention.

FIG. 7 is a block circuit diagram showing a schematic configuration of a noise reduction device of a third exemplary embodiment of the present invention.

FIG. 8 is a flowchart for explaining learning in the third embodiment of the present invention.

FIG. 9 is a block circuit diagram showing a schematic configuration of a noise reduction device according to a fourth exemplary embodiment of the present invention.

FIG. 10 is a flow chart for explaining learning in the fourth embodiment of the present invention.

FIG. 11 is a block circuit diagram showing a schematic configuration of a noise reduction device according to a fifth exemplary embodiment of the present invention.

FIG. 12 is a block circuit diagram showing a schematic configuration of a noise reduction device of a sixth exemplary embodiment of the present invention.

FIG. 13 is a block circuit diagram showing a schematic configuration of a noise reduction device of a seventh exemplary embodiment of the present invention.

FIG. 14 is a block circuit diagram showing a schematic configuration of a noise reduction device according to an eighth exemplary embodiment of the present invention.

FIG. 15 is a block circuit diagram showing a schematic configuration of a conventional noise reduction device.

[Explanation of symbols]

2, 7, 13, 32, 52, 66, 82, 106 Blocking circuit 3, 8, 16, 38, 54, 68, 90, 112 Class code generating circuit 4, 9, 17, 40, 56, 70, 92 , 114 RO
M table 10 block decomposition circuit 14, 34, 84, 108 filter

─────────────────────────────────────────────────── ─── Continuation of front page (72) Inventor Takeharu Nishikata 6-735 Kita-Shinagawa, Shinagawa-ku, Tokyo Within Sony Corporation (56) Reference JP-A-4-42682 (JP, A) JP-A 3-285461 (JP, A) JP-A-4-8070 (JP, A) JP-A-5-167892 (JP, A) JP-A-5-137026 (JP, A) JP-A-4-207292 (JP, A) (58) Fields investigated (Int.Cl. ⁷ , DB name) H04N 5/14-5/217 G06T 5/00

Claims (18)

(57) [Claims] 1. A noise reduction device for outputting an image signal in which noise is reduced compared to an input image signal, wherein a blocking means for dividing the input image signal into a plurality of blocks and block data are used. Class classification means for classifying each block into a plurality of classes by pattern classification and outputting a class code corresponding to the classified class; and corresponding to the block of the input image signal according to the class code of each block. A noise reduction device, comprising: a noise reduction image signal output means for outputting a noise-reduced image signal. 2. The noise reduced image signal output means has a storage means for storing noise reduced low noise image data corresponding to a class code obtained by classifying blocks of an image containing noise. 2. The noise according to claim 1, wherein the storage means is adapted to read the low noise image data with the class code generated by the class classification means as an address in order to generate the noise reduced image signal. Reduction device. 3. A class code of a block obtained by blocking an image signal containing noise using an image signal containing noise and a corresponding low noise image signal, and image data of a block corresponding to the low noise image signal. Based on the image data of the low-noise image signal corresponding to the same class code, representative low-noise image data is obtained in advance, and the storage means uses the class code as an address for the representative low-noise image data. 3. The noise reduction apparatus according to claim 2, wherein the low noise image data is stored. 4. The low-noise image data is pixel data, the storage means stores corresponding pixel data using each class code as an address, and the class code generated by the class classification means is used as an address. 4. The noise reduction device according to claim 3, wherein low noise image data is output by reading pixel data. 5. The low noise image data is pixel data in block units, and the storage means stores pixel data in block units corresponding to each class code as an address, and is generated by the class classification means. 4. The noise reduction device according to claim 3, wherein low noise image data is output by reading the pixel data in block units with a class code as an address. 6. The noise reduction device according to claim 5, further comprising a conversion unit for converting the pixel data of the low noise block unit into a noise reduced image signal arranged in the order of image signal lines. 7. The noise reduction device according to claim 1, wherein the class classification means performs class classification processing on the noise reduction filter output of the input image signal. 8. A filter means for filtering the blocked image signal from the blocking means, and a difference between the blocked image signal from the blocking means and the output signal from the filter means. Difference output means, the class classification means performs class classification processing on the difference from the difference output means to generate a class code, and the noise reduced image signal output means is responsive to the class code. 2. The noise reduction device according to claim 1, further comprising difference data generation means for outputting difference data of the noise-reduced image data using the difference from the difference output means. 9. The noise-reduced image signal output means includes an addition means for adding an output signal from the difference data generation means and an output signal from the filter means. Noise reduction device. 10. The noise reduced image signal output means comprises:
Storage means for storing coefficient data corresponding to a class code obtained by classifying blocks of an image containing noise, and coefficient data read from the storage means into a corresponding value of the block of the input image signal. Noise reduction image data generation means for multiplying to obtain low noise image data, and the storage means uses the class code generated by the class classification means as an address to generate the noise reduction image signal. The noise reduction device according to claim 1, wherein coefficient data is read out. 11. A filter means for filtering the blocked image signal from the blocking means, and a difference between the blocked image signal from the blocking means and the output signal from the filter means. A difference output unit, the class classification unit performs a class classification process on the difference from the difference output unit to generate a class code, and the noise reduced image signal output unit includes a noise-containing image Storage means for storing coefficient data corresponding to a class code obtained by classifying blocks, and coefficient data read from the storage means is multiplied by a corresponding value of the block of the input image signal to reduce noise. The noise according to claim 1, further comprising difference data generation means for obtaining difference data between the image data and an image containing noise. Reduction device. 12. The noise reduced image signal output means comprises:
The noise reducing device according to claim 11, further comprising an adding unit that adds an output signal from the difference data generating unit and an output signal from the filtering unit. 13. A noise reduction method for outputting an image signal in which noise is reduced compared to an input image signal, wherein a blocking step of dividing the input image signal into a plurality of blocks and blocked data are used. A class classification step of classifying each block into a plurality of classes by pattern classification and outputting a class code corresponding to the classified class; and corresponding to the block of the input image signal according to the class code of each block. And a noise reduction image signal output step of outputting a noise-reduced image signal. 14. The noise-reduced image signal output step uses storage means for storing noise-reduced low-noise image data corresponding to a class code obtained by classifying blocks of an image containing noise. 14. The noise reduction according to claim 13, wherein the storage means reads the low noise image data by using the class code generated by the class classification means as an address in order to generate the noise reduced image signal. Method. 15. A class code of a block obtained by blocking an image signal containing noise using an image signal containing noise and a corresponding low noise image signal, and image data of a block corresponding to the low noise image signal. Based on the image data of the low-noise image signal corresponding to the same class code, representative low-noise image data is obtained in advance, and the storage means uses the class code as an address for the representative low-noise image data. 15. The noise reduction method according to claim 14, wherein the low noise image data is stored. 16. The noise reduction method according to claim 13, wherein the class classification step classifies the noise reduction filter output of the input image signal. 17. The noise reduction method according to claim 13, wherein the class classification step performs class classification on the difference between the input image signal and the output of the noise reduction filter. 18. The noise-reduced image signal output step uses storage means for storing coefficient data corresponding to a class code obtained by classifying blocks of an image containing noise, and is read from the storage means. And a noise-reduced image data generating step of multiplying the corresponding value of the block of the input image signal by a corresponding value of the block of the input image signal to obtain low-noise image data. 14. The noise reduction method according to claim 13, wherein the coefficient data is read using the class code generated by the class classification means as an address.