JP5372586B2 - Image processing device - Google Patents

Description

The present invention relates to an image processing apparatus that determines whether a specific area of an image is an edge portion or a flat portion and improves image quality.

  Conventionally, noise reduction processing, band enhancement processing, and the like are known as methods for improving the image quality of a photographic image acquired by a digital camera or the like. For these processes, in order to achieve both noise reduction and sharpness improvement, methods such as applying different processes to the flat area and edge area in the image or changing processing parameters are generally used. Yes.

For example, Patent Document 1 discloses a high image quality technique that switches between a smoothing process suitable for an edge part and a smoothing process suitable for a flat part according to the result of the process of determining the edge part and the flat part. ing. Japanese Patent Application Laid-Open No. 2004-228620 further discriminates an edge portion or a flat portion and further calculates pixel isolation, and determines sharpness processing conditions based on the edge portion / flat portion discrimination result and the isolation to improve image quality. A technique is disclosed. In this case, the discrimination between the edge portion and the flat portion in these conventional examples includes, as a discrimination index, edge strength, local dispersion, local connectivity in the vicinity of the point of interest, color correlation and directionality between color channels in the case of a color image, etc. Is used.
Japanese Patent No. 3477603 Japanese Patent Laid-Open No. 11-331600

  However, in the edge / flat portion discrimination method in the above-described conventional example, there is a high possibility that the discrimination index itself is affected by noise when the amount of noise is large, and the edge portion and the flat portion cannot be accurately discriminated. is there. In addition, if the effect of noise included in the final determination index is reduced by performing discrimination after smoothing the image in the vicinity of the target pixel, or by smoothing the determination index itself, a texture portion including fine edges There is also a problem of misclassifying as a flat part.

  The present invention has been made in view of the above circumstances, and an object of the present invention is to provide an image processing apparatus that can reliably determine a flat portion without being affected by noise and can improve image quality.

The invention according to claim 1 is between a statistic relating to a pixel value of an image obtained by imaging in advance using an imaging system for acquiring a digital image and a parameter value relating to a probability distribution of noise in the image. Data holding means for holding the relationship as noise characteristic data , means for calculating a predetermined statistic from a distribution of pixel values in a region including the target pixel in the digital image, the noise characteristic data and the predetermined statistic from the amount, and calculates the parameter calculating means for calculating a value of a parameter related to the probability distribution of the noise in the region including the pixel of interest, the probability distribution of said noise by using the value of the parameter, the pixels in the region including the pixel of interest the distribution of values, and the probability distribution of the noise which is calculated using the values of the parameters, how matching A test means for calculating an index value indicating the Luke, is characterized by comprising a.

  The invention according to claim 2 is characterized in that, in the invention according to claim 1, the parameter relating to the probability distribution of noise is standard deviation or variance.

According to a third aspect of the present invention, in the first aspect of the present invention, the statistic is an average or median of pixel values in a region including the target pixel.

According to a fourth aspect of the present invention, in the first aspect of the invention, the noise characteristic data gives a standard deviation or variance of noise in a flat portion having the input value as a pixel value with respect to a predetermined input value. , and the said parameter calculating means, the statistic as the input value for the noise characteristic data, the standard deviation or variance on the probability distribution of the noise in the region including the said target pixel as a feature to seek the value of the parameter Yes.

  The invention according to claim 5 is the invention according to claim 1, characterized in that the test means is a test method for evaluating the similarity of the cumulative distribution function of the probability distribution.

According to a sixth aspect of the invention, in the first aspect of the invention, the distribution of pixel values in the region including the target pixel is a distribution of pixel values of pixels at discrete positions in the region including the target pixel. It is characterized by that.

Calculating invention of claim 7, in the invention according to the first aspect, pixel values definitive to each pixel of the digital image, by linear computation of pixel values of an image obtained by imaging using the imaging system It is characterized by comprising a new pixel value component.

  According to the present invention, since the flat portion can be accurately determined without being affected by noise, the edge portion and the flat portion can be reliably determined, and a high-quality image with edge enhancement can be obtained. .

  Further, according to the present invention, since a texture portion that is easily affected by noise can be determined without being mistaken as a flat portion, a high-quality image in which the texture portion is emphasized can be acquired.

  Furthermore, according to the present invention, it is possible to acquire a high-quality image in which an edge portion and a texture portion are emphasized for a color image captured by an imaging system.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

(First embodiment)
FIG. 1 shows a schematic configuration of a digital camera to which the image processing apparatus according to the first embodiment of the present invention is applied.

  In FIG. 1, reference numeral 100 denotes a digital camera. The digital camera 100 includes an imaging system 101, an input image buffer 102, a noise model table 103 as data holding means, an edge flatness determination unit 104, a noise reduction unit 105, and an edge enhancement unit 106. And a recording unit 107.

  The imaging system 101 includes a monochrome imaging device, converts an optical image of a subject into a digital image, performs optical black correction, and outputs the digital image to the input image buffer 102. The input image buffer 102 stores the digital image output from the imaging system 101 for each pixel (x, y). The noise model table 103 records, as noise characteristic data, a data pair of an average pixel value and a standard deviation of a patch image for an image obtained when images of various brightness are captured in advance by the imaging system 101. Details of the noise model table 103 will be described later. The edge flatness determination unit 104 sequentially scans each pixel (x, y) in the input image buffer 102 to determine whether the vicinity of the pixel (x, y) is flat or edge-like. The edge index E (x, y) that becomes smaller at is calculated and output. Details of the edge flatness determination unit 104 will be described later. When the edge index E (x, y) is input from the edge flatness determination unit 104, the noise reduction unit 105 selects a neighboring pixel (a pixel in a predetermined range around the corresponding pixel) at the corresponding pixel position (x, y). Used to perform a known edge-preserving smoothing process for noise reduction. The edge enhancement unit 106 includes an internal buffer (not shown), and performs edge enhancement processing by accumulating all the pixel values after noise reduction input from the noise reduction unit 105 in the internal buffer. The recording unit 107 includes a recording medium, and records the edge enhancement result acquired by the processing in the edge enhancement unit 106.

  FIG. 2 shows a schematic configuration of the edge flatness determination unit 104, which includes a neighborhood buffer 110, a statistic calculation unit 111, a variance calculation unit 112 as a parameter calculation unit, and a normality test unit 113 as a test unit. Yes.

  The neighborhood buffer 110 temporarily stores a pixel value in the vicinity of the pixel at coordinates (x, y) read from the input image buffer 102, that is, a pixel value of the target pixel and a pixel value in a predetermined range around the target pixel. The statistic calculator 111 obtains a median A from the pixel values of neighboring pixels stored in the neighboring buffer 110. The variance calculation unit 112 refers to the noise characteristic data in the noise model table 103 and obtains the noise amount Noise (A) (standard deviation σ) corresponding to the median A as a parameter related to the normal distribution of noise. The normality test unit 113 obtains the distribution (observation data) of the neighboring pixel values in the neighborhood buffer 110 from the median value A obtained by the statistic calculation unit 111 and the standard deviation σ obtained by the variance calculation unit 112. It is determined whether the noise is close to the normal distribution N (A, σ). There are various known methods for determining whether the observation data matches the normal distribution. Here, the kormogolov-smirnov method is used.

  Next, the operation of the digital camera 100 configured as described above will be described.

  Now, when a shutter (not shown) of the digital camera 100 is pressed, the imaging system 101 converts the optical image of the subject into a digital image, performs optical black correction, and outputs the digital image to the input image buffer 102. The input image buffer 102 stores the digital image output from the imaging system 101 for each pixel (x, y).

  Next, the edge flatness determination unit 104 sequentially scans each pixel (x, y) in the input image buffer 102 and determines whether the vicinity of the pixel (x, y) is flat or edge-like. Then, an edge index E (x, y) that is large at the edge portion and small at the flatness based on the determination result is calculated, and the calculated edge index E (x, y) is calculated as the noise reduction unit 105 and the edge enhancement unit. It outputs to 106.

  In this case, the edge flatness determination unit 104 obtains the edge index E as follows. First, the pixel value near the pixel at the coordinate (x, y) is read from the input image buffer 102 and stored in the neighborhood buffer 110. Next, the statistic calculation unit 111 obtains the median A as a statistic regarding the pixel values of the neighboring pixels stored in the neighboring buffer 110. Further, the variance calculation unit 112 refers to the noise model table 103 to obtain a noise amount Noise (A) (standard deviation σ) corresponding to the median value A. More specifically, it is generally known that noise generated in an image pickup device has a normal distribution, and there is a device-specific relationship as shown in FIG. 6 between the standard deviation and the pixel value, for example. Therefore, in the noise model table 103, a data pair of the pixel value average (median value) and standard deviation of the patch image is recorded for an image obtained when images of various brightness are captured in advance by the imaging system 101. deep. Each black dot in FIG. 6 corresponds to these. In the variance calculation unit 112, the noise amount Noise (A) corresponding to the median A, that is, the standard deviation σ is obtained by linear interpolation by the polygonal line function Noise (x) determined at each black point in FIG.

Finally, in the normality test unit 113, the distribution (observation data) of the pixel values of the neighboring pixels in the neighborhood buffer 110 is the median value A obtained by the statistic calculation unit 111 and the standard obtained by the variance calculation unit 112. It is determined using a known kormogolov-smirnov method whether it is close to (or matches) the normal distribution N (A, σ) of noise having a deviation σ. In this method, the cumulative frequency distribution of the observation data is calculated, and the maximum value of the difference between the cumulative frequency distribution and the normal distribution cumulative distribution function is used as a normality determination index (edge index E). In this embodiment, as shown in FIG. 3, the cumulative histogram h (v) of the pixel values of the neighboring pixels in the neighboring buffer 110 and the cumulative distribution function L * S () of the noise normal distribution N (A, σ) The difference between (vA) / σ) is taken and the maximum value of this difference is used as the edge index E. Then, the edge index E is output to the noise reduction unit 105 and the edge enhancement unit 106. In this case, the edge index E is obtained by the following equation.

E = max (| h (v) -L * S ((vA) / σ) |) mm <= v <= MM
Where S (x) is the cumulative distribution function of normal distribution N (0,1) with A = 0 and σ = standard deviation 1, mm and MM are the minimum and maximum pixel values in the neighborhood buffer 110, and L is This is the number of pixels in the neighborhood buffer 110.

  The edge index E thus obtained for the edge flatness determination unit 104 is output to the noise reduction unit 105 and the edge enhancement unit 106.

  In the noise reduction unit 105, when the edge index E (x, y) is input from the edge flatness determination unit 104, a known edge-preserving smoothing process is performed using a neighboring pixel at the corresponding pixel position (x, y). Do. Specifically, the pixel value U (x, y) after noise reduction is calculated by the following formula 1.

U (x, y) = ΣΣ_ij wij * V (x + i, y + j) / ΣΣwij, (-N <= i, j <= N, N is the neighborhood size) (Equation 1)
wij = W (| V (x + i, y + j) -V (x, y) |, σ)
W (x, σ) = exp (-x * x / (σ * σ))
In this case, since the weight function W (x, σ) increases as x decreases, the neighboring pixel V (x + i, y + j) whose value is close to V (x, y) is heavier according to the calculation of Equation 1. An average is taken by weighting. As a result, the edges are averaged using only the pixels on the same edge as the target pixel in the edge part, and the edges are maintained. In the flat part, all the neighboring pixels are averaged with an equal weight. Will be smoothed. Further, σ in the equation is controlled as shown in FIG. 4 based on the edge index E input from the edge flatness determination unit 104. In this case, when σ is small, the weighting function W (x, σ) takes a small value even for a large x, and the change in the value of the weighting function relative to the change in the value of x becomes relatively large, whereas σ is If the value is large, the weighting function changes relatively little even if the value of x slightly changes, and generally has a characteristic that takes a large value. Therefore, in the edge portion where the edge index E is large as shown in FIG. By increasing σ in the flat part where σ is small and the edge index E is small, weight uniformity increases in the flat part regardless of the difference between the pixel values of the target pixel and neighboring pixels, and the smoothing effect increases. In the edge portion, smoothing is not applied, and the edge rounding can be reduced.

  The pixel value U (x, y) after noise reduction at the pixel (x, y) obtained by the calculation in the noise reduction unit 105 is output to the edge enhancement unit 106.

  The edge enhancement unit 106 accumulates the noise-reduced pixel values input from the noise reduction unit 105 in an internal buffer (not shown). Then, when all of the nearby noise reduction pixel values necessary for the edge enhancement process at the pixel (x, y) are accumulated in the internal buffer, the edge enhancement process is performed. Specifically, the pixel value U (x + i, y + j) after noise reduction in the vicinity of the pixel (x, y) (−M <= i, j <= M, M is the neighborhood size) is K Edge components are extracted by (x, y) = ΣΣ_ij f (i, j) * U (x + i, y + j), and an edge enhancement result T (x, y) is calculated by the following equation.

T (x, y) = U (x, y) + α (E) * K (x, y)
Here, f (i, j) is a bandpass filter for edge extraction, and α is an edge enhancement amount shown in FIG. In this case, the edge enhancement amount α is small when the edge index E is small and is large when the edge index E is large, so that the noise is weakened more in the flat portion and the edge is more enhanced in the edge portion.

  The edge enhancement result T (x, y) calculated by the edge enhancement unit 106 is output to the recording unit 107 and recorded on the recording medium. Thereafter, when a series of processing is completed for all pixels (x, y), the digital camera 100 ends the imaging operation and enters a standby state.

  As described above, in this embodiment, a data pair of the pixel value average (median value) and standard deviation of a patch image is obtained for an image obtained by previously capturing patches of various brightnesses by an imaging system. It is recorded in the noise model table 103 as characteristic data, and a standard deviation σ corresponding to the median value A of neighboring pixel values of the input digital image acquired from the imaging system is estimated from the noise characteristic data, and this standard deviation σ And the normal distribution N (A, σ) of the noise having the median A is determined whether the distribution of the pixel values of the neighboring pixels is close, and the noise reduction unit 105 determines whether or not the distribution of the pixel values of the neighboring pixels is close. The noise enhancement and edge enhancement by the edge enhancement unit 106 are executed. As a result, the flat portion can be accurately determined without being affected by noise, so that the edge portion and the flat portion can be reliably determined, and a high-quality image with edge enhancement can be obtained. . In addition, since a texture portion that is easily affected by noise can be determined without being mistaken as a flat portion, a high-quality image in which the texture portion is emphasized can be acquired.

  Various modifications can be assumed for the above-described embodiment. For example, the neighborhood data read to the neighborhood buffer 110 may be an area having any shape around the coordinates (x, y), and may be arranged discretely. In this case, the edge flatness determination in the edge flatness determination unit 104 is performed on the read shape region or the region covered by the discrete points.

  Further, as a method for controlling the noise reduction processing and the edge enhancement processing using the edge index E, any known method for discriminating and processing the edge portion and the flat portion can be used.

Furthermore, in the above-described embodiment , an example of a monochrome imaging system has been described, but the same processing can be applied to a three-plate imaging system. In this case, the edge flatness determination may be processed by regarding each of the R component, the G component, and the B component as independent monochrome images, or a determination result obtained by integrating the edge flatness determination results may be generated. As an integration method, a method used in any known method for performing edge flatness determination using information on a plurality of channels for a color image can be used.

Further, it is assumed a normal distribution as a distribution of the form in the noise of the embodiment described above, is also applicable to noise with any cumulative distribution function. In that case, the appropriate cumulative distribution function may be replaced with the normal distribution cumulative distribution function, and kormogolov-smirnov or any known distribution test method may be used. Further, in the above-described embodiment, the noise normal distribution is used as the probability distribution, but other probability distributions such as a Poisson distribution may be used. Furthermore, in the above description, the standard deviation σ has been described as a parameter relating to the normal distribution.

(Second embodiment)
Next, a second embodiment of the present invention will be described.

  FIG. 7 shows a schematic configuration of a digital camera to which the image processing apparatus according to the second embodiment of the present invention is applied. The same parts as those in FIG.

In FIG. 7, reference numeral 200 denotes a digital camera. The digital camera 200 includes a single-plate imaging system 201, an input image buffer 202, a noise model table 103, an edge flatness determination unit 204, a synchronization unit 205, a color conversion unit 206, a gradation. A conversion unit 207, a YC conversion unit 208, a noise reduction unit 209, an edge enhancement unit 210, a compression unit 211, and a recording unit 107 are included .

The single-plate imaging system 201 includes a single-plate imaging system including a primary color Bayer imaging device, converts an optical image of a subject into a digital image, and performs optical black correction and white balance correction. The input image buffer 202 stores the image after white balance for each pixel (x, y). The edge flatness determination unit 204 sequentially scans each pixel (x, y) in the input image buffer 202, and the edge indices Ey (x, y), Ecr (x, y), Ecb at each pixel (x, y). Calculate (x, y). Details of the edge flatness determination unit 204 will be described later. The synchronization unit 205 performs a known interpolation process at each pixel position (x, y), and R (x, y), G (x, y), B (x, y) that are RGB components at each pixel position. ) Is generated. The color conversion unit 206 applies a 3 × 3 color conversion matrix to R (x, y), G (x, y), and B (x, y) input from the synchronization unit 205, and applies The results R ′ (x, y), G ′ (x, y), and B ′ (x, y) are output. The gradation conversion unit 207 outputs a predetermined gradation conversion function γ to the results R ′ (x, y), G ′ (x, y), B ′ (x, y) after the color conversion in the color conversion unit 206. Perform gradation conversion by , and output the result. The YC conversion unit 208 calculates the luminance component Y (x, y) and the color difference components Cr (x, y), Cb (x, y) from the result after the gradation conversion by the gradation conversion unit 207 by a known method. And output this result.

The noise reduction unit 209 corresponds to the edge index Ey (x, y) corresponding to the luminance component Y (x, y) and the color difference components Cr (x, y), Cb (x, y) input from the YC conversion unit 208. ), Ecr (x, y), Ecb (x, y) are obtained and noise reduction processing is performed. Edge enhancement unit 210, it outputs without processing for the chrominance components input from the noise reduction unit 209, only the luminance component, performs the same processing as in the first embodiment described above. When the final luminance component is input from the edge enhancement unit 210, the compression unit 211 performs a known compression process such as JPEG, and outputs the result to the recording unit 107.

  FIG. 8 shows a schematic configuration of the edge flatness determination unit 204. The neighborhood buffer 220, the RGB statistic calculation unit 221, the RG statistic calculation unit 227, the BG statistic calculation unit 228, the G variance calculation unit 222, and the RG variance The calculation unit 223, the BG variance calculation unit 224, the RG generation unit 225, the BG generation unit 226, the RG normality test unit 229, the BG normality test unit 230, and the G normality test unit 231 are configured.

The neighborhood buffer 220 temporarily stores pixel values near the pixel at the coordinates (x, y) read from the input image buffer 202. The RGB statistic calculation unit 221 calculates a statistic for each color component for the pixel values in the neighborhood buffer 220. In this embodiment, the median Ar for each color component as a statistic, Ag, seeking Ab. The RG statistic calculation unit 227 and the BG statistic calculation unit 228 obtain the median value Arg of the RG value and the median value Abg of the BG value in the neighborhood buffer 220 as statistics for the RG and BG components in the neighborhood buffer 220. The G variance calculation unit 222 refers to the noise model table 103 and obtains the standard deviation σg, which is the noise amount Ng corresponding to the median Ag, as a parameter related to the normal distribution of noise. The RG variance calculating unit 223 similarly refers to the noise model table 103 to obtain the noise amount Nr of the R component corresponding to the median Ar, and the standard deviation σrg of the noise of the color difference RG component near the coordinates (x, y). Calculate Similarly, the BG variance calculating unit 224 calculates the standard deviation σbg of the noise of the color difference BG component in the vicinity of the coordinates (x, y).

  The R-G generation unit 225 and the B-G generation unit 226 calculate the R-G and B-G components at the R and B pixel positions in the neighborhood buffer 220 after the processing of the RGB statistic calculation unit 221 is completed. The G normality test unit 231, the RG normality test unit 229 and the BG normality test unit 230 respectively test whether the pixel value distributions of the G component, RG component, and BG component in the neighborhood buffer 220 match the normal distribution. I do. For this test method, an Anderson-Darling method different from the first embodiment described above is used.

  Next, the operation of the digital camera 200 configured as described above will be described.

  Now, when a shutter (not shown) of the digital camera 200 is pressed, the single-plate imaging system 201 converts the optical image of the subject into a digital image, performs optical black correction and white balance correction, and then enters the input image buffer 202. Output. In this case, three color components are required to obtain a color image, but an image input to the input image buffer 202 can obtain only one type of color component per pixel as shown in FIG. 9A. Therefore, the white balance correction at this time is performed by applying the white balance coefficient Wr to the pixel position where R is obtained and applying the white balance coefficient Wb to the pixel position where B is obtained.

When an image after white balance correction is obtained in the input image buffer 202, the synchronization unit 205 performs a known interpolation process for each pixel position (x, y), and R (x, y), G (x, y), and B (x, y) are generated and output to the color conversion unit 206. The color conversion unit 206 applies a 3 × 3 color conversion matrix to the input R (x, y), G (x, y), B (x, y), and the result R ′ (x, y), G ′ (x, y), and B ′ (x, y) are output to the gradation conversion unit 207. The gradation conversion unit 207 applies gradation conversion using a predetermined gradation conversion function γ to the results R ′ (x, y), G ′ (x, y), and B ′ (x, y) after color conversion. The result shown in the following expression is output to the YC conversion unit 208.

R '' (x, y) = γ (R '(x, y)), G''(x, y) = γ (G' (x, y)), B '' (x, y) = γ (B '(x, y)),
The YC conversion unit 208 calculates the luminance component Y (x, y) and the color difference components Cr (x, y), Cb (x, y) from the result after the gradation conversion by the gradation conversion unit 207 by a known method. Then, the result shown in the following expression is output to the noise reduction unit 209.

Y (x, y) = 0.3 * R`` (x, y) + 0.6 * G '' (x, y) + 0.1 * B '' (x, y)
Cr (x, y) = R`` (x, y) -Y (x, y)
Cb (x, y) = B`` (x, y) -Y (x, y)
On the other hand, in parallel with the processing so far, the edge flatness determination unit 204 has edge indices Ey (x, y), Ecr () of the luminance component and the color difference component in the vicinity of each pixel (x, y) of the input image buffer 202. x, y), Ecb (x, y) are calculated and sequentially output to the noise reduction unit 209 and the edge enhancement unit 210.

  In this case, the edge flatness determination unit 204 obtains an edge index as follows. In this case, the edge flatness determination unit 204 reads out the pixel value near the pixel at the coordinate (x, y) from the input image buffer 202 to the neighborhood buffer 220. In this case, the data read to the neighborhood buffer 220 is, for example, as shown in FIG.

  Next, the median values Ar, Ag, and Ab are obtained by the RGB statistic calculation unit 221 as the statistic for each color component for the pixel values in the neighborhood buffer 220. In this state, the G variance calculation unit 222 refers to the noise model table 103 to obtain a noise amount Ng (standard deviation σg) corresponding to the median value Ag. Similarly, the RG variance calculating unit 223 also refers to the noise model table 103 to obtain the noise amount Nr of the R component corresponding to the median Ar, and the noise of the color difference RG component near the coordinates (x, y). The standard deviation σrg is calculated from the following.

Nr = Wr * Noise (Nr / Wr)
σrg = (Nr * Nr + Ng * Ng) ^ (1/2)
Here, the function Noise (x) is a polygonal line function defined by the data in the noise model table 103, and reflects the noise characteristics of the single-plate imaging system 201 as in the first embodiment described above. It is. In calculating the noise amount Nr of the R component, since the white balance is corrected by the single-plate imaging system 201 and the coefficient Wr is applied, calculation for correcting the influence is performed. Further, in the calculation of the noise amount Nrg of the RG component, the standard deviation σrg is calculated from the standard deviation σr of the R component and the standard deviation σg of the G component using the additivity of normal distribution.

The BG variance calculation unit 224 performs the same calculation to calculate the standard deviation σbg of the noise of the color difference BG component in the vicinity of the coordinates (x, y). In this case, Wb is used instead of Wr as the white balance coefficient, and the noise standard deviation of the B component is calculated in place of the noise standard deviation σr of the R component.

When the processing of the RGB statistic calculation unit 221 is completed, the RG generation unit 225 and the BG generation unit 226 perform parallel processing with the G variance calculation unit 222 in the neighborhood buffer 220 as shown in FIG. RG and BG components are calculated at the R pixel and B pixel positions. In this case, at each R pixel position and B pixel position, one of four G pixels around is selected and the differences RG and BG are calculated. In either any which G to select always may choose the right G, as shown in the shaded area in FIG. 9 (a), may be changed position to select a G for each pixel. In general, the direction in which G is selected is not uniform in the neighborhood buffer 220, and it is desirable not to select a common G between adjacent R pixels or B pixels.

  When the RG generation unit 225 and the BG generation unit 226 calculate the RG and BG components in the neighborhood buffer 220, the RG statistics calculation unit 227 and the BG statistics calculation unit 228 calculate the statistics of the RG and BG components in the neighborhood buffer 220. As the quantities, the median value Arg of the RG values and the median value Abg of the BG values in the neighboring buffer 220 are obtained.

  When the processing so far is completed, the standard deviations σg, σrg, σbg obtained from the noise characteristics of the single-plate imaging system 201 and the median values Ag, Arg, Abg are obtained for each of the G component, the RG component, and the BG component. Next, in the G normality test unit 231, the RG normality test unit 229, and the BG normality test unit 230, whether the pixel value distribution of the G component, the RG component, and the BG component in the neighborhood buffer 220 matches the normal distribution, respectively. Perform a test of whether or not. Unlike the first embodiment described above, the Anderson-Darling method is used as the test method here. In this Anderson-Darling method, an index A2 indicating whether the data string v1, v2,... VN matches the probability distribution f (x) is calculated as follows.

  i) Rearrange v1 to vN in ascending order and let u1 to uN.

ii) Calculate S = Σ (2k-1) * (log (F (uk) + log (F (uN + 1-k)) / N (sum is 1 <= k <= N)
iii) A2 = -NS
Here, log (x) is a natural logarithm of x, and F (x) is a cumulative distribution function of the probability distribution f (x).

  Based on this method, the G normality test unit 231 uses the data component v1 to vN as values of the G component obtained in the neighborhood buffer 220 to determine whether or not the normal distribution N (Ag, σg) is satisfied. ) To iii), and A2 of the result is set as the edge index Ey. As the cumulative distribution function, the cumulative distribution function S ((x−Ag) / σg) of the normal distribution N (Ag, σg) is used. (Here, S is a cumulative distribution function of a normal distribution with mean 0 and standard deviation 1.) Similarly, the RG normality test unit 229 uses the value of the RG component obtained in the neighborhood buffer 220 as the above data. Whether columns v1 to vN match the normal distribution N (Arg, σrg) is calculated by i) to iii), and A2 of this result is set as the edge index Erg. Further, the BG normality test unit 230 determines whether or not the values of the BG components obtained in the neighborhood buffer 220 conform to the normal distribution N (Abg, σbg) as the above data strings v1 to vN according to i) to iii). The A2 of this result is used as the edge index Ebg.

In the noise reduction unit 209, the G normality test unit 231 and the RG normality are applied to the luminance component Y (x, y) and the color difference components Cr (x, y), Cb (x, y) input from the YC conversion unit 208. Noise reduction processing is performed at the stage where the edge indices Ey (x, y), ECr (x, y), ECb (x, y) input from the sex test unit 229 and the BG normality test unit 230 are given, The result is output to the edge enhancement unit 210. Here, as the noise reduction processing is performed in the same manner as edge preserving smoothing as described in the first embodiment, but the brightness, and applied separately to each color difference, the lower also the value of the standard deviation σ different luminance and color difference It is expressed by a formula.

Y '(x, y) = ΣΣ_ij wij * Y (x + i, y + j) / ΣΣwij, (-N <= i, j <= N, N is the neighborhood size)
wij = W (| Y (x + i, y + j) -Y (x, y) |, σy)
Ck '(x, y) = ΣΣ_ij wij * Ck (x + i, y + j) / ΣΣwij,
wij = W (| Ck (x + i, y + j) -Ck (x, y) |, σck)
(-N <= i, j <= N, N is the neighborhood size, k is r or b)
W (x, σ) = exp (-x * x / (σ * σ))
Here, σy and σc are functions βy and βc shown in FIG. 10, and are calculated from the edge indices Ey and Ec according to the following equations.

σy = βy (Ey (x, y)), σcr = βc (Ecr (x, y)), σcb = βc (Ecb (x, y))
Here, the functions βy and βc have a relationship in which βc is larger when the edge index is small, and βc is smaller when the edge index is larger. As a result, extreme smoothing and extreme edge preservation for edges with strong contrast work for the color difference component, and processing more favorable as noise reduction characteristics for the color difference component can be realized.

The edge enhancement unit 210, as it is output without processing for the inputted color difference component, only the luminance component, and outputs by performing the same processing as in the first embodiment. When the final luminance component and color difference component are input from the edge enhancement unit 210, the compression unit 211 performs a known compression process such as JPEG, and the result is output to the recording unit 107 and recorded on the recording medium. The

  Even in this case, as in the first embodiment, the flat portion can be accurately determined without being affected by noise in the color image picked up by the image pickup system. It is possible to reliably determine the quality of the image, to obtain a high-quality image with edge enhancement, and to distinguish a texture portion that is easily affected by noise without being misidentified as a flat portion. Images can be acquired.

  Various modifications can be assumed for the second embodiment. For example, FIG. 11 shows a simplified process of the edge flatness determination unit 204, and the same parts as those in FIG. In this case, instead of the RG variance calculation unit 223 and the BG variance calculation unit 224 shown in FIG. 8, an R variance calculation unit 241 and a B variance calculation unit 242 are provided, and the RG normality test unit 229 and the BG normality test unit 230 are provided. Instead, an R normality test unit 243 and a B normality test unit 244 are provided, and instead of performing edge flatness determination for the color differences RG and BG, edge flatness determination of the R component and B component is performed. Thus, a common edge index is obtained for the common luminance component and color difference component.

  In this configuration, the R variance calculation unit 241 obtains the standard deviation σr of the R component noise, and the R normality test unit 243 converts the R component obtained in the neighborhood buffer 220 into the normal distribution N (Ar, σr). It is tested whether or not they match, and the index A2 is obtained as the edge index Er of the R component. Similarly, the B variance calculation unit 242 obtains the standard deviation σb of the B component noise, and the B normality test unit 244 matches the B component obtained in the neighborhood buffer 220 with the normal distribution N (Ab, σb). The index A2 is obtained as the edge index Eb of the B component. Finally, the maximum value calculation unit 245 obtains the maximum values Emax of the edge indices Eg, Er, and Eb of the G component, R component, and B component, and the edge index Ey of the luminance component and the edge indices Ecr, Ecb of the color difference components are all represented by Emax. Is output to the noise reduction unit 209 and the edge enhancement unit 210.

  In addition, this invention is not limited to the said embodiment, In the implementation stage, it can change variously in the range which does not change the summary.

  Furthermore, the above embodiments include inventions at various stages, and various inventions can be extracted by appropriately combining a plurality of disclosed constituent elements. For example, even if some constituent requirements are deleted from all the constituent requirements shown in the embodiment, the problem described in the column of the problem to be solved by the invention can be solved, and is described in the column of the effect of the invention. If the above effect is obtained, a configuration from which this configuration requirement is deleted can be extracted as an invention.

1 is a diagram showing a schematic configuration of a digital camera to which an image processing apparatus according to a first embodiment of the present invention is applied. The figure which shows schematic structure of the edge flat determination part used for 1st Embodiment. The figure explaining how to obtain | require the edge flatness determination parameter | index in 1st Embodiment. The figure explaining the method of calculating | requiring the weighting function in the noise reduction part used for 1st Embodiment according to an edge parameter | index. The figure explaining the method of calculating | requiring the edge enhancement amount in the edge enhancement part used for 1st Embodiment according to an edge parameter | index. The figure explaining the example of data in the noise model table used for 1st Embodiment. The figure which shows schematic structure of the digital camera to which the image processing apparatus concerning the 2nd Embodiment of this invention is applied. The figure which shows schematic structure of the edge flat determination part used for 2nd Embodiment. The figure explaining the example of data in the neighborhood buffer used for 2nd Embodiment. The figure explaining the method of calculating | requiring the weighting function in the noise reduction part used for 2nd Embodiment according to an edge parameter | index. The figure which shows schematic structure of the modification of the edge flat determination part used for 2nd Embodiment.

DESCRIPTION OF SYMBOLS 100, 200 ... Digital camera, 101, 201 ... Imaging system 102 ... Input image buffer, 103 ... Noise model table 104 ... Edge flatness determination part, 105 ... Noise reduction part 106 ... Edge emphasis part, 107 ... Recording part, 110 ... Neighborhood Buffer 111 ... Statistics calculation unit 112 ... Distribution calculation unit 113 ... Normality test unit 202 ... Input image buffer 204 ... Edge flatness determination unit 205 ... Synchronization unit 206 ... Color conversion unit 207 ... Tone conversion unit 208 ... YC conversion unit, 209 ... noise reduction unit 210 ... edge enhancement unit, 211 ... compression unit, 220 ... neighboring buffer 221 ... RGB statistic calculation unit, 222 ... R variance calculation unit 223 ... RG variance calculation unit, 224 ... BG Variance calculation unit, 225 ... RG generation unit 226 ... BG generation unit, 227 ... RG statistic calculation unit 228 ... BG statistic calculation unit, 229 ... RG normality test 230 ... BG Normality Test section, 231 ... G Normality Test unit 241 ... R variance calculation unit, 242 ... B variance calculation unit, 243 ... R Normality Test unit 244 ... B Normality Test section, 245 ... maximum value calculator

Claims (7)

The relationship between the statistic relating to the pixel value of the image obtained by imaging in advance using an imaging system for acquiring a digital image and the parameter value relating to the probability distribution of noise in the image is defined as noise characteristic data. Data holding means for holding;
Means for calculating a predetermined statistic from a distribution of pixel values in a region including the target pixel in the digital image;
Parameter calculation means for calculating a value of a parameter relating to a probability distribution of noise in a region including the target pixel from the noise characteristic data and the predetermined statistic;
Using the values of the parameters to calculate the probability distribution of the noise, the the distribution of pixel values in a region including the pixel of interest, and the probability distribution of the noise which is calculated using the values of the parameters, how matching A test means for calculating an index value indicating whether to perform ,
An image processing apparatus comprising:   The image processing apparatus according to claim 1, wherein the parameter relating to the probability distribution of noise is standard deviation or variance. The image processing apparatus according to claim 1, wherein the statistic is an average or median of pixel values in a region including the target pixel. The noise characteristic data gives a standard deviation or variance of noise in a flat portion having the input value as a pixel value with respect to a predetermined input value, and the parameter calculation means calculates the statistic as the noise characteristic. as input values for the data, the image processing apparatus according to claim 1, wherein the determining of the standard deviation or variance for the probability distribution of the noise in the region including the said target pixel as the value of the parameter.   The image processing apparatus according to claim 1, wherein the test unit is a test method for evaluating a similarity of a cumulative distribution function of a probability distribution. The distribution of the pixel values in the region including the target pixel, the image processing apparatus according to claim 1, characterized in that the distribution of the pixel values of the pixels of the discrete positions in the area including the pixel of interest. Pixel values definitive to each pixel of the digital image, claims, characterized in that it consists of a new pixel value component calculated by the linear operation of the pixel values of the image obtained by imaging using the imaging system The image processing apparatus according to 1 .

Images