JP4399449B2 - Image noise reduction system, image noise reduction method, and computer program - Google Patents

Description

  The present invention relates to an image noise reduction system, an image processing method, and a computer program, and more particularly to a deblocking filter.

  As one of the processes for reducing noise superimposed on an image, a low-pass filter (hereinafter referred to as “LPF”) process using edge information (contour information) is known. If the detection accuracy of the edge information used in the filter processing is low, the LPF processing is performed even on pixels that should not be subjected to the LPF processing, and a good processing result cannot be obtained. On the other hand, if the detection accuracy of the edge information is increased, it is possible to obtain a good processing result, but on the other hand, the amount of calculation required for the processing increases, and the processing time becomes long and the circuit scale increases. Occurs.

  Conventionally, a filter called a deblocking filter is also known. A deblocking filter is a filter that reduces the distortion of the block boundary that occurs in the encoded characters of an image. H., which is a moving picture compression technique adopted in the MPEG-4 standard. In H.264, this deblocking filter is incorporated in the encoding loop as an in-loop filter. An advantage of using this filter is that an image from which block noise has been removed can be used as a reference image. By removing block noise in the reference image, it is possible to remove the influence of noise from the prediction error due to motion compensation prediction, so that the coding efficiency can be improved.

  Although not related to image noise reduction processing, Patent Document 1 discloses an image processing apparatus that suppresses overshoot and undershoot and emphasizes a fine structure portion (high-frequency component or detail structure) of an original image. ing. This image processing apparatus has a low-frequency image creation unit, a difference unit, and an addition unit. The low frequency image creation unit creates a low frequency image of the original image. The difference unit creates a difference image by subtracting the low frequency image from the original image. The adding unit adds an amount corresponding to the absolute value of the density value of the difference image to the original image using a predetermined filter function. Thereby, the fine structure portion of the original image is emphasized.

  Further, although it relates to image compression rather than image noise reduction processing, Patent Document 2 discloses AC component prediction.

JP 2000-293683 A JP 2004-289290 A

  As image noise reduction processing, processing using wavelet transform is known in addition to the above-described LPF processing. However, when a simple base is used, it is difficult to effectively reduce block noise, which is high-frequency component noise, because a low-frequency component is included in the AC component of the filter. In addition, when a complex base is used, there are inconveniences such as an increase in time required for inner product calculation and an increase in circuit scale.

  The present invention has been made in view of such circumstances, and an object thereof is to provide a novel deblocking filter that can effectively reduce noise superimposed on an original image with a relatively small amount of calculation. .

In order to solve such a problem, the first invention is an image that is recursively executed while reducing the size of a pixel area to be subjected to noise reduction processing of an original image in accordance with a decrease in k (k is a natural number). A noise reduction system is provided. This system includes a reduction processing unit, an interpolation processing unit, a subtraction unit, a noise reduction unit, and an addition unit. In the noise reduction processing in which k is the maximum value, the reduction processing unit divides the original image into pixel regions having a pixel size of 2 ^k × 2 ^k , and calculates a first pixel representative value for each pixel region. While the first reduced image is generated, in the noise reduction processing in which k is not the maximum value, the processed image fed back as the output of the (k + 1) noise reduction processing is used as the first reduced image. In addition, the reduction processing unit divides the original image into pixel regions having a pixel size of 2 ^k−1 × 2 ^k−1 regardless of k, and calculates a second pixel representative value for each pixel region. A second reduced image is generated. The interpolation processing unit enlarges the first reduced image by the interpolation process using the first pixel representative value, thereby increasing the third pixel for each pixel area having a pixel size of 2 ^k−1 × 2 ^k−1. A low frequency image composed of representative values is generated. The subtracting unit subtracts the third pixel representative value of each pixel area of the low-frequency image from the second pixel representative value of each pixel area of the second reduced image, so that the first difference consisting of the high-frequency component is subtracted. Generate an image. The noise reduction unit compares a pixel value constituting the first difference image with a threshold value to perform a noise reduction process for changing the pixel value, and generates a second difference image as an output of the noise reduction process To do. The adding unit adds the low-frequency image generated by the interpolation processing unit and the second difference image generated by the noise reduction unit, thereby generating a processed image in which noise of the original image is reduced. Then, the processed image generated by the noise reduction process in which k is the minimum value is output as a final output image. Also. The processed image generated by the noise reduction process where k is not the minimum value is fed back as the input of the noise reduction process of (k−1).

The second invention provides an image noise reduction method for recursively executing a noise reduction process of an original image while reducing the size of a pixel area to be processed according to a decrease in k (k is a natural number). In this method, in the noise reduction processing in which k is the maximum value, the original image is 2 ^k × 2 ^k In the noise reduction processing in which k is not the maximum value, the first reduced image is generated by dividing the pixel area into pixel areas of the pixel size and calculating the first pixel representative value for each pixel area. The processed image fed back as the output of the noise reduction processing is used as the first reduced image, and the original image is converted to 2 regardless of k. ^k-1 × 2 ^k-1 A first step of generating a second reduced image by dividing the pixel area into pixel areas of the pixel size and calculating a second pixel representative value for each pixel area, and interpolation using the first pixel representative value By processing, the first reduced image is enlarged by 2 ^k-1 × 2 ^k-1 From the second step of generating a low-frequency image composed of the third pixel representative value for each pixel area of the pixel size, and from the second pixel representative value of each pixel area of the second reduced image, A third step of generating a first difference image composed of high-frequency components by subtracting the third pixel representative value of each pixel region, and a pixel value constituting the first difference image as a threshold value A noise reduction process for changing the pixel value in comparison with the fourth step of generating a second difference image as an output of the noise reduction process, a low-frequency image, and a second difference image. By adding together, a fifth step of generating a processed image in which noise of the original image is reduced, and a processed image generated by the noise reduction processing in which k is the minimum value are output as a final output image, and k Is reduced by noise reduction processing that is not the minimum value. Processing images, and a sixth step of feedback as an input of the noise reduction processing of (k-1).

According to a third aspect of the invention, a computer is caused to execute an image noise reduction method that is recursively executed while reducing the size of a pixel area to be subjected to noise reduction processing of an original image in accordance with a decrease in k (k is a natural number). Provide a computer program. This program reduces the original image to 2 in the noise reduction process where k is the maximum value. ^k × 2 ^k In the noise reduction processing in which k is not the maximum value, the first reduced image is generated by dividing the pixel area into pixel areas of the pixel size and calculating the first pixel representative value for each pixel area. The processed image fed back as the output of the noise reduction processing is used as the first reduced image, and the original image is converted to 2 regardless of k. ^k-1 × 2 ^k-1 A first step of generating a second reduced image by dividing the pixel area into pixel areas of the pixel size and calculating a second pixel representative value for each pixel area, and interpolation using the first pixel representative value By processing, the first reduced image is enlarged by 2 ^k-1 × 2 ^k-1 From the second step of generating a low-frequency image composed of the third pixel representative value for each pixel area of the pixel size, and from the second pixel representative value of each pixel area of the second reduced image, A third step of generating a first difference image composed of high-frequency components by subtracting the third pixel representative value of each pixel region, and a pixel value constituting the first difference image as a threshold value A noise reduction process for changing the pixel value in comparison with the fourth step of generating a second difference image as an output of the noise reduction process, a low-frequency image, and a second difference image. By adding together, a fifth step of generating a processed image in which noise of the original image is reduced, and a processed image generated by the noise reduction processing in which k is the minimum value are output as a final output image, and k Is reduced by noise reduction processing that is not the minimum value. The processed image is to execute the image noise reduction method and a sixth step of feedback as an input of the noise reduction processing of (k-1) to the computer.

According to the present invention, the noise reduction processing for changing the pixel value by comparing the pixel value constituting the first difference image with the threshold value, the size of the pixel area is reduced by k (k is a natural number). The noise superimposed on the original image can be more effectively reduced by executing recursively while reducing the size.

(First embodiment)
FIG. 1 is a configuration diagram of an image noise reduction system according to the present embodiment. This image noise reduction system is a kind of so-called deblocking filter that reduces noise superimposed on an original image. The image noise reduction system includes a low-frequency image generation unit 1, a subtraction unit 2, a noise reduction unit 3, an addition unit 4, and a filter unit 5.

  The low frequency image generation unit 1 generates a low frequency image of the original image, that is, an image composed of low frequency components of the original image. In the present embodiment, the low-frequency image generation unit 1 includes two functional blocks, a reduction processing unit 1a and an interpolation processing unit 1b. A plurality of pixel regions are defined in advance by dividing an original image of one frame vertically and horizontally on the image plane. For each pixel region, the reduction processing unit 1a calculates a single pixel representative value that represents the pixel region. As the pixel representative value, typically, the pixel average value (DC component) in the pixel region to be processed can be used. As a result of collecting a plurality of pixel values in the pixel area into pixel representative values (one value), an image defined as a set of pixel representative values becomes a reduced image in which the size of the original image is reduced.

  FIG. 2 is an explanatory diagram of a reduced image as an example. The size of one frame of the original image is m × n pixels, and the size of one pixel region S is 2 × 2 pixels. In this case, the pixel representative value of the pixel area S is the four pixels (i, j), (i + 1, j), (i, j + 1), (i + 1, j + 1) pixel average value. A reduced image that is a set of pixel representative values has a size obtained by halving the size of the original image vertically and horizontally, that is, (m / 2) × (n / 2) pixels. Note that the pixel representative value only needs to represent the characteristic of the pixel value existing in the pixel region S by some value. Therefore, in addition to the above-described pixel averaging method, so-called thinning processing or the like may be used. For example, a predetermined value in the pixel region S, for example, the pixel value at the upper left (i, j) is used as the pixel representative value of the pixel region S.

  The interpolation processing unit 1b performs an interpolation process on the reduced image generated by the reduction processing unit 1a, and expands the reduced image to the same size (m × n pixels) as the original image, thereby generating a low-frequency image. By this interpolation processing, one pixel of the reduced image is enlarged to a size corresponding to the pixel block S of the original image. When one pixel of a reduced image is interpolated / enlarged, the pixel representative value (one value) of this one pixel and the pixel representative values (multiple values) of peripheral pixels located around this pixel are used. In the present embodiment, AC component prediction is used as an example of interpolation processing.

  FIG. 3 is an explanatory diagram of AC component prediction. For one pixel (processing target) in the reduced image, four pixels (i, j), (i + 1, j), (i, j + 1), (i + 1, j + 1) are obtained by AC component prediction. Pixel values for (that is, the same size as the pixel block S of the original image) are calculated. Regarding the input value, the central pixel value is C, the pixel value immediately above is U, the pixel value just below is D, the pixel value right is R, and the pixel value right is L. Further, regarding the output value, the interpolation pixel value of the upper left sub-pixel (i, j) in the processing target is a, the interpolation pixel value of the lower left sub-pixel (i + 1, j) is b, and the upper right sub-pixel (i , j + 1) is the interpolation pixel value c, and the interpolation pixel value of the lower right sub-pixel (i + 1, j + 1) is d. In this case, each pixel value a to d is calculated by the following equation.

(Calculation formula of pixel values a to d)
a = (4 × C + 2 × L + 2 × U) / 8
b = (4 × C + 2 × R + 2 × U) / 8
c = (4 × C + 2 × L + 2 × D) / 8
d = (4 × C + 2 × R + 2 × D) / 8

  That is, a certain interpolated pixel value (for example, a) is obtained with respect to the pixel representative value C of the pixel to which its own subpixel (i, j) belongs and the pixel representative values L and U of two pixels adjacent to its own subpixel. In addition, after adding predetermined weights (4/8, 2/8, 2/8), these are added. By subdividing one pixel into four sub-pixels, the reduced image is eventually returned to the same size as the original image and output as a low-frequency image. A low-frequency image generated by interpolation of adjacent values is an image composed of low-frequency components. Therefore, visually, the low-frequency image expresses edges (locations where changes between adjacent pixels are large) more smoothly than the original image, and in particular, locations with minute changes (superimposed on the original image). (Including noise spots) is also smoothed.

  When AC component prediction is used as the interpolation processing, there is an advantage that the circuit configuration can be simplified. This is because most of the multiplications and divisions performed in the AC component prediction process are powers of 2, so that these operations can be executed by bit shift using shifters instead of arithmetic units such as multipliers and dividers.

  Various methods other than the above-described methods are known for generating a low-frequency image of an original image, and any method may be adopted. Below, three methods (1)-(3) are illustrated as a typical thing.

(1) Recursive AC component prediction Recursive AC component prediction generates a low-frequency image by adding the frequency component of the pixel area in the reduced image and the frequency component of the original image corresponding to the pixel area in a positional manner. It is a technique to do. Specifically, first, the frequency component of the pixel region in the reduced image is obtained by the following equation. Similar to FIG. 3, when five pixel values C, U, D, L, and R are defined, three frequency components α ′, β ′, and γ ′ are obtained based on the following expression for the pixel region to be processed. Is calculated.

(Calculation formula of frequency components α ', β', γ ')
α ′ = (L−R + 1) / 4
β ′ = (U−D + 1) / 4
γ ′ = 0

  Next, frequency components α ″, β ″, γ ″ of the original image corresponding in position to the pixel area to be processed are calculated. Pixel values relating to the positions (i, j), (i + 1, j), (i, j + 1), (i + 1, j + 1) in the pixel region to be processed are (a ′, b ′). , C ′, d ′), the three frequency components α ′, β ′, g are calculated by the following equations.

(Calculation formula of frequency components α ″, β ″, γ ″)
α ″ = (a′−b ′ + c′−d ′) / 2−α ′
β ″ = (a ′ + b′−c′−d ′) / 2−β ′
γ ″ = (a′−b′−c ′ + d ′) / 2

  Then, as shown in the following expression, the frequency components α, β, γ are calculated by adding the two frequency components (for example, α ′, α ″) described above for each frequency component.

(Calculation formula of frequency components α, β, γ)
α = α '+ α''
β = β '+ β''
γ = γ '+ γ''

  Finally, these frequency components are converted to obtain the pixel values of the pixels constituting the target pixel region. The pixel values (a, b, c, d) in the pixel area to be processed are calculated from the following equation.

(Calculation formula of pixel values a to d)
a = C + (α + β + γ)
b = C + (β−γ−α)
c = C + (α−β−γ)
d = C + (γ−α−β)

  Similar to the case of the AC component prediction described above, there is an advantage that the circuit configuration can be simplified even when the recursive AC component prediction is used. This is because all multiplications and divisions performed in the process of recursive alternating current component prediction are powers of 2, so that the arithmetic unit can be configured mainly with a shifter. In particular, when the frequency component is known in advance, the amount of calculation is smaller than that of the AC component prediction, so that the processing speed can be increased. In this case, it is necessary to input the pixel value of the original image to the interpolation processing unit 1b (in FIG. 1, an arrow enters from the “original image” to the “interpolation processing unit 1b”).

(2) Linear interpolation After connecting pixels with a straight line, a pixel value is calculated by estimating a pixel on the straight line by linear approximation.

(3) Convolution interpolation (cubic convolution)
Also called a cubic interpolation method, a pixel to be estimated is obtained by applying a convolution function to pixel values of pixels at surrounding grid points. As a convolution function, a sic function (f (x) = sinπx / πx) having a property that the value decreases as the distance from the origin increases.

  Note that the above-described method using two processes of original image reduction and interpolation enlargement is theoretically equivalent to a method of directly performing low-pass filter processing on the original image. Therefore, instead of the above-described method, a low-pass filter may be directly applied to the original image. However, in comparison with the low-pass filter processing, the above-described method has an advantage that the arithmetic expression (polynomial) can be simplified.

  The subtraction unit 2 generates a first difference image by calculating a difference between the original image and the low frequency image generated by the low frequency image generation unit 1 for each pixel. Specifically, the pixel value of the low frequency image corresponding to this position is subtracted from the pixel value of the original image. Thereby, the first difference image is an image composed of high-frequency components. This is because if a low frequency image consisting only of low frequency components is subtracted from the original image, only the high frequency components remain. Therefore, as a visual feature, an image in which an edge or a minute change portion (including a noise portion) is extracted is obtained.

  The noise reduction unit 3 performs a noise reduction process for each pixel using a pixel value constituting the first difference image as an input value, and generates a second difference image constituted by the output value of the noise reduction process. Of the input values constituting the first difference image, those having a small value are regarded as noise. In general, the difference between an edge and high-frequency noise is that the brightness difference (amplitude) between adjacent pixels tends to be smaller in the latter than in the former. Therefore, if the threshold value is appropriately set in consideration of this tendency, an edge (a high-frequency component having a large amplitude) and noise (a high-frequency component having a small amplitude) can be appropriately separated depending on the magnitude of the input value. For the input value regarded as noise, at least the output value is made smaller (reduced) than the input value.

  FIG. 4 is an input / output characteristic diagram as a first example in the noise reduction processing. The input value x is the pixel value of the first difference image, and the output value f (x) is the pixel value of the second difference image. When the input value x is less than or equal to the threshold value T, 0 is output as the output value f (x). On the other hand, when the input value x is larger than the threshold value T, the input value x is output as it is as the output value f (x). In the region of x> T, the output value f (x) also increases linearly as the input value x increases. This input / output characteristic can also be expressed by a change amount (that is, a slope) defined as a change in output value f (x) with respect to a change in input value x. Assuming that the amount of change in the region x <T is A1 (= 0), the amount of change A2 (= 1) in the region x> T satisfies the relationship 1 ≧ A2> A1. Further, the change amount A3 (= ∞) in the region near x = T satisfies the relationship of A3> A2.

  FIG. 5 is an input / output characteristic diagram as a second example in the noise reduction processing. When the input value x is less than or equal to the threshold value T, approximately 0 is output as the output value f (x). On the other hand, when the input value x is larger than the threshold value T, the input value x is output almost as it is as the output value f (x). In the region of x> T, the output value f (x) increases continuously as the input value x increases. Further, when this input / output characteristic is expressed by the above-described change amount, the change amount (not constant) in the region of x <T is A1≈0. The amount of change (not constant) in the region of x> T satisfies the relationship 1 ≧ A2> A1. Further, the amount of change (not constant) in the region near x = T satisfies the relationship of A3> A2.

  As another threshold processing, the input value x may be quantized / inversely quantized by the quantization coefficient T. In this case, when the quantization step is set to T and the input / output characteristics are increased stepwise, the final output value f (x) becomes 0 in the region where the input value x is 0 to T. This corresponds to threshold processing using the quantization step T as a threshold.

  The addition unit 4 generates a processed image by adding the low-frequency image generated by the low-frequency image generation unit 1 and the second difference image generated by the noise reduction unit 3 for each pixel. Specifically, the pixel value of the low-frequency image and the pixel value of the second difference image corresponding to the position are added. As described above, the low frequency image is composed only of the low frequency component. Therefore, if the low-frequency image is added with the second differential image (noise reduction processed) consisting only of the high-frequency component, an image including the low-frequency component and the high-frequency component other than noise (mainly the edge) is processed. Generated as an image. When compared with the original image, the processed image has an image quality superior to that of the original image by reducing the noise superimposed on the original image.

  The filter unit 5 uses the second difference image generated by the noise reduction unit 3 as edge information, and performs a low-pass filter process on the processed image generated by the addition unit 4 to generate an output image. That is, an edge in an image is detected based on a second difference image (a high-frequency image mainly composed of an edge), and LPF processing is performed only on a region that is not an edge portion (applicable LPF processing). In consideration of the possibility that block noise is included in the low-frequency image, noise filter processing for reducing block noise may be performed together.

  According to the present embodiment, it is possible to realize a deblocking filter that can provide visually good image quality. The original image includes a low frequency component and a high frequency component due to edges and noise. By subtracting the difference between the original image and the low frequency image in the subtracting unit 2, the low frequency component of the original image is removed, and the high frequency component of the original image is extracted as the first difference image. The first difference image includes a high frequency component caused by an edge and a high frequency component caused by noise. Therefore, the noise reduction unit 3 performs noise reduction processing on the first difference image to remove the high frequency component of the noise and extract the high frequency component of the edge as the second difference image. Then, the addition unit 4 adds the low-frequency image and the second difference image, thereby adding the low-frequency component and the high-frequency component of the edge. As a result, a processed image obtained by removing high frequency components of noise from the original image is obtained. The processed image output from the adder 4 has a visually good image quality compared to the original image. In addition, since the amount of computation required for processing is small, it is possible to perform high-speed processing enough to follow real-time playback of moving images, and to realize with a relatively small circuit scale when implemented in hardware. Can do.

  In addition, as in the present embodiment, if the applied LPF process is further performed on the processed image subjected to the noise reduction process, the image quality of the output image can be further improved. In this adaptive LPF process, the second difference image calculated in the pre-processing process is used as edge information. Therefore, even if an additional LPF process is added, the calculation amount does not need to be increased so much.

  In the present embodiment, reduction processing and subsequent interpolation processing based on alternating current component prediction are performed in order to generate a low-frequency image. The multiplication and division necessary for these processes can basically be realized by a bit shift process. Therefore, the filter process using the frequency domain such as the wavelet transform can be performed substantially while using the spatial domain, so that it is possible to realize a highly accurate deblocking process while maintaining a relatively small circuit scale. .

(Second Embodiment)
In the first embodiment described above, the low-frequency image output from the interpolation processing unit 1b includes up to half the frequency component of the original image, but cannot include even smaller frequency components than this frequency component. Therefore, there is a possibility that block noise is included in the low-frequency image itself, and as a result, it may be a factor that hinders improvement in the image quality of the output image. In the present embodiment, the image quality is further improved by recursively repeating the noise reduction process.

  FIG. 9 is a configuration diagram of an image noise reduction system according to the second embodiment. The structural feature of this system is that a determination unit 6 is provided on the output side of the filter 5 and is output as a final output image according to the determination result, or the previous stage (that is, interpolation) as a transient image. It is a point which switches whether it feeds back to the input side of the process part 1b. Along with this, some modifications have been made to the functions of the reduction processing unit 1a and the subtraction unit. Since the other points are the same as those in the first embodiment described above, the same reference numerals are given and description thereof is omitted.

The reduction processing unit 1a generates a first reduced image and a second reduced image based on the original image, and supplies the former to the interpolation processing unit 1b and the latter to the subtraction unit 2, respectively. In order to generate the first reduced image, first, the original image is divided into (2 ^k ) × (2 ^k ) (k is a natural number) pixel regions. Then, similarly to the above-described first embodiment, the pixel representative value (first pixel representative value) of each pixel region is calculated. The first reduced image is defined as a set of first pixel representative values. As a result, the size of the first reduced image is 1/2 ^{k of the} original image on one side. On the other hand, in order to generate the second reduced image, first, the original image is divided into (2 ^k−1 ) × (2 ^k−1 ) pixel regions. Then, similarly to the first embodiment described above, the pixel representative value (second pixel representative value) of each pixel region is calculated, and the second reduced image is defined as a set of second pixel representative values. The If the size of the original image is m × n, the size of the second reduced image is 1/2 ^{k of the} original image on one side.

The interpolation processing unit 1b performs an interpolation process on the first reduced image, and expands the first reduced image to the same size as the second reduced image, thereby generating a low-frequency image. Interpolation processing is performed by AC component prediction as in the first embodiment. The generated low-frequency image is divided into (2 ^k−1 ) × (2 ^k−1 ) pixel areas, and pixel representative values (third pixel representative values) are calculated in the respective pixel areas. . Therefore, the size of the low-frequency image is 1/2 ^{k−1 on} one side of the original image.

The subtraction unit 2 outputs a difference between the low-frequency image generated by the interpolation processing unit 1b and the second reduced image generated by the reduction processing unit 1a as a first difference image. Specifically, the third pixel representative value of the low-frequency image corresponding to this position is subtracted from the second pixel representative value of each pixel region. At this time, since the low-frequency image has the same size as the second reduced image (one side is 1/2 ^{k-1 of the} original image), the representative pixels corresponding to each other can be subtracted.

The determination unit 6 manages the loop variable k used in the processing units 1a and 1b, and determines whether to output as a final output image based on the variable k. The initial value of the loop variable that defines the number of repetitions of the noise reduction process is set in advance, and is decremented by one as the process progresses. Until the loop variable k reaches 1 (k> 1), the output image from the filter unit 5 is fed back to the interpolation processing unit 1b as a transient image (k ← k−1). On the other hand, when the loop variable k reaches 1 (k = 1), the output image from the filter unit 5 is output as the final output image. One side of the output image from the filter unit 5 is 1/2 ^k-1 of the original image. This is the final output image when k = 1, that is, when it has the same size as the original image. In other words, as long as the size of the output image is smaller than that of the original image, it is fed back as the first reduced image.

  FIG. 7 is an explanatory diagram of image noise reduction processing according to the second embodiment. As an example, a case will be described in which the initial value of the loop variable k is 3, that is, the noise reduction process is repeated three times.

  When k = 3, the reduction processing unit 1a generates a 1/8 reduced image whose one side is 1/8 of the original image and a 1/4 reduced image whose one side is 1/4 of the original image. The A 1/4 low frequency image is generated by the interpolation processing unit 1b. Then, the subtracting unit 2 subtracts a 1/4 low-frequency image of the same size from the 1/4 reduced image. Thereafter, a ¼ output image is generated by the filter unit 5 through the same processing as in the first embodiment described above. Since k ≠ 1, the determination unit 6 feeds back the 1/4 output image to the interpolation processing unit 1b, and the loop variable k is set to 2.

  When k = 2, the reduction processing unit 1a generates a 1/2 reduced image whose one side is 1/2 of the original image. On the other hand, a 1/4 output image fed back by the determination unit 7 is used as the 1/4 reduced image. A 1/2 low frequency image is generated by the interpolation processing unit 1b. The subtracting unit 2 subtracts a ½ low-frequency image having the same size from the ½ reduced image. Finally, a 1/2 output image is generated by the filter unit 5. Since k ≠ 1, the determination unit 6 feeds back the 1/2 output image to the interpolation processing unit 1b and sets the loop variable k to 1.

  When k = 1, a 1/1 reduced image (original image itself) having the same size as the original image and a 1/2 output image fed back by the determination unit 7 are used. A 1/1 low frequency image is generated by the interpolation processing unit 1b. The subtracting unit 2 subtracts a 1/1 low-frequency image having the same size from the 1/1 reduced image. Finally, a 1/1 output image is generated by the filter unit 5. Since k = 1, a 1/1 output image is output as the final output of the determination unit 6.

  According to the present embodiment, the output image generated from the filter unit 5 corresponding to the original image is fed back to the interpolation processing unit 1b, and the noise reduction processing is repeated a plurality of times. By such a process, a low frequency image having a plurality of types of low frequency components is recursively applied, so that noise superimposed on the original image can be more effectively compared with the first embodiment described above. Can be reduced.

  In the above-described embodiments, the configuration and operation of the noise reduction system have been mainly described. However, a computer program that causes a computer to execute such a function also constitutes another embodiment of the present invention.

1 is a configuration diagram of an image noise reduction system according to a first embodiment. Illustration of reduced image as an example Illustration of AC component prediction Input / output characteristic diagram as a first example in noise reduction processing Input / output characteristic diagram as a second example in noise reduction processing Configuration of Image Noise Reduction System According to Second Embodiment [ Explanatory drawing of the image noise reduction process concerning 2nd Embodiment.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Low frequency image generation part 1a Reduction processing part 1b Interpolation processing part 2 Subtraction part 3 Noise reduction part 4 Addition part 5 Filter part 6 Determination part

Claims (3)

In an image noise reduction system that recursively executes a noise reduction process of an original image while reducing the size of a pixel area to be processed according to a decrease in k (k is a natural number) ,
In the noise reduction processing in which k is the maximum value, the original image is divided into pixel areas having a pixel size of 2 ^k × 2 ^{k, and} a first pixel representative value for each pixel area is calculated to obtain a first pixel value. While generating a reduced image, in the noise reduction process where k is not the maximum value, the processed image fed back as the output of the noise reduction process of (k + 1) is used as the first reduced image, and regardless of the k, A reduction processing unit that generates a second reduced image by dividing the original image into pixel regions having a pixel size of 2 ^k-1 × 2 ^k-1 and calculating a second pixel representative value for each pixel region. When,
By enlarging the first reduced image by the interpolation process using the first pixel representative value , the third pixel representative value for each pixel region having a pixel size of 2 ^k−1 × 2 ^k−1 is obtained. An interpolation processing unit for generating a low-frequency image,
By subtracting the third pixel representative value of each pixel region of the low-frequency image from the second pixel representative value of each pixel region of the second reduced image, the first consisting of a high-frequency component A subtracting unit that generates a difference image and outputs ;
A noise reduction unit that compares a pixel value constituting the first difference image with a threshold value and performs a noise reduction process for changing the pixel value, and generates a second difference image as an output of the noise reduction process When,
Addition for generating the processed image in which noise of the original image is reduced by adding the low-frequency image generated by the interpolation processing unit and the second difference image generated by the noise reduction unit And
The processed image generated by the noise reduction process in which k is the minimum value is output as a final output image,
The image noise reduction system , wherein the processed image generated by the noise reduction process in which k is not the minimum value is fed back as an input of the noise reduction process of (k-1) . In an image noise reduction method for recursively executing a noise reduction process of an original image while reducing the size of a pixel area to be processed according to a decrease in k (k is a natural number) ,
In the noise reduction processing in which k is the maximum value, the original image is divided into pixel areas having a pixel size of 2 ^k × 2 ^{k, and} a first pixel representative value for each pixel area is calculated to obtain a first pixel value. While generating a reduced image, in the noise reduction process where k is not the maximum value, the processed image fed back as the output of the noise reduction process of (k + 1) is used as the first reduced image, and regardless of the k, A ^first reduced image is generated by dividing the original image into pixel areas having a pixel size of 2 ^k-1 × 2 ^k-1 and calculating a second pixel representative value for each pixel area. Steps,
By enlarging the first reduced image by the interpolation process using the first pixel representative value , the third pixel representative value for each pixel region having a pixel size of 2 ^k−1 × 2 ^k−1 is obtained. A second step of generating a low frequency image comprising:
By subtracting the third pixel representative value of each pixel region of the low-frequency image from the second pixel representative value of each pixel region of the second reduced image, the first consisting of a high-frequency component A third step of outputting to generate a difference image;
A noise reduction process for changing the pixel value by comparing a pixel value constituting the first difference image with a threshold value, and generating a second difference image as an output of the noise reduction process. Steps,
A fifth step of generating the processed image in which noise of the original image is reduced by adding the low-frequency image and the second difference image;
The processed image generated by the noise reduction processing in which k is the minimum value is output as a final output image, and the processed image generated in the noise reduction processing in which k is not the minimum value is (k−1). And a sixth step of feeding back as an input of the noise reduction processing in (2) . In a computer program for causing a computer to execute an image noise reduction method that is recursively executed while reducing the size of a pixel area to be processed for noise reduction processing of an original image in accordance with a decrease in k (k is a natural number) ,
In the noise reduction processing in which k is the maximum value, the original image is divided into pixel areas having a pixel size of 2 ^k × 2 ^{k, and} a first pixel representative value for each pixel area is calculated to obtain a first pixel value. While generating a reduced image, in the noise reduction process where k is not the maximum value, the processed image fed back as the output of the noise reduction process of (k + 1) is used as the first reduced image, and regardless of the k, A ^first reduced image is generated by dividing the original image into pixel areas having a pixel size of 2 ^k-1 × 2 ^k-1 and calculating a second pixel representative value for each pixel area. Steps,
By enlarging the first reduced image by the interpolation process using the first pixel representative value , the third pixel representative value for each pixel region having a pixel size of 2 ^k−1 × 2 ^k−1 is obtained. A second step of generating a low frequency image comprising:
By subtracting the third pixel representative value of each pixel region of the low-frequency image from the second pixel representative value of each pixel region of the second reduced image, the first consisting of a high-frequency component A third step of outputting to generate a difference image;
A noise reduction process for changing the pixel value by comparing a pixel value constituting the first difference image with a threshold value, and generating a second difference image as an output of the noise reduction process. Steps,
A fifth step of generating the processed image in which noise of the original image is reduced by adding the low-frequency image and the second difference image;
The processed image generated by the noise reduction processing in which k is the minimum value is output as a final output image, and the processed image generated in the noise reduction processing in which k is not the minimum value is (k−1). A sixth step of feeding back as an input of the noise reduction processing of
A computer program for causing a computer to execute an image noise reduction method including:

Images