Method of reducing block and mosquito noise (effect) in images

0001The present invention relates to a method of reducing block noise as well as mosquito noise in an image, which noises are caused at the time of decoding encoded, compressed image data on a block-by-block.

0002In order to produce a highly compressed image data, an orthogonally transformation encoding as a highly effective image data compression technique is generally employed. This technique in summary involves dividing an input image signal into blocks of such as 8×8 pixels by a blocking circuit, subjecting the divided blocks to an orthogonal transformation process by an orthogonal. transformation circuit to generate a frequency component signal, subjecting the orthogonally transformed data to a linear quantization process by a quantization circuit using a predetermined quantization step width, and allocating a variable length code to the result of the quantisation by a variable-longth encoding circuit to generate an encoded image signal. This orthogonal transform includes such as discrete Fourier transform, Walsh-Hadamard transform, Karhunen-Loevel transform and discrete cosine transform (DCT). Of them, the DCT is most widely used.

0003On the other hand, in a decoding apparatus for regenerating an image signal corresponding to the input image signal from the encoded image signal generated by the above compression encoding apparatus, the encoded image signal is subjected to variable length decoding by a variable-length decoding circuit, and reverse quantization is performed by a reverse-quantization circuit using a predetermined quantization step width and the image signal is regenerate by reverse orthogonal transform.

0004Herein, the orthogonal transform and the quantization which are both nonreversible transform contain errors in the regenerated image signal obtained by the decoding apparatus. Particularly, quantization errors in quantization and reverse quantization deteriorate the quality of the regenerated image. The largei the quantization step width (or the larger the compression rate), the greater the number of quantization errors caused and hence the more noticeable the deterioration of the quality of the regenerated image signal. Image deterioration inherent to such orthogonal transform include block noise and mosquito noise, the former resulting from discontinuity along the boundary of adjacent blocks which is perceived like a mosaic appearance, the latter being perceived like swarms of mosquitoes clustered around a contour of a character or a figure on the background of the image (hereinafter referred simply to "a contour").

0005In order to reduce block noise, Japanese Patent Application Laid-open, No. 1998-191335 (Prior Reference 1) proposes a technique that involves detecting boundaries of the blocks and filtering the pixels on the boundaries, thus correcting an image. In order to reduce mosquito noise, Japanese Patent Application Laid-open No. 1998-164576 (Prior Reference 2) proposes a technique that involves taking a difference between a target pixel and its surrounding pixels such as ±3 pixels, and applying a certain weighing coefficient thereto and adding the result to the target value when the absolute value of the difference is smaller than a threshold value.

0006The block noise reducing method as proposed in the Prior Reference 1 may be incapable of determining whether an object concerned is a boundary causing a great difference in pixel value between adjacent pixels due to block noise generated or only a contour of a character or a figure in image having a large difference in pixel value between adjacent pixels. If the contour in the image is erroneously determined as a boundary of the blocks and accordingly subjected to a block noise reducing process, the contour in the image may be blurred, which leads to deterioration of the image quality.

0007The mosquito noise reducing method as proposed in the Prior Reference 2 is effective since it is a simple filtering process and therefore, takes only a short time for the processing; on the other hand, this simple filtering process is hard to properly reduce mosquito noise without deterioration of the image quality because it only involves a somewhat rough process, which may also lead to blurred contour and hence deterioration of the image quality.

0008As long as quantization errors are inevitably caused by quantization and reverse quantization, both block noise and mosquito noise appear in relevant regions of a regenerated image, whether they are great or not. In spite of this fact, there are no techniques, which have been found up to the present, to satisfactorily reduce both block noise and mosquito noise.

0009It is an object of the present invention to provide a method of reducing noise in images that is capable of properly reducing both block noise and mosquito noise without deterioration of the image quality.

0010According to the present invention, there is provided a method of reducing block noise and mosquito noise in an image, in which block noise and mosquito noise are caused at the time of decoding encoded, compressed image data on a block-by-block basis. The method includes: dividing each of luminance component image data and color-difference component image data of the image data into blocks corresponding to the blocks created in the encoding and decoding of the image data; applying a first filtering step to the luminance component image data with each of pixels on the boundaries of the blocks being designated as a target pixel, thereby creating a first luminance component image data with the boundaries of the blocks smoothened; performing a second filtering step with each pixel of the first luminance component image data being designated as a target value, thereby creating a second luminance component image data with its entirety smoothened; creating edge image data by subtracting each pixel value of the second luminance component image data from a corresponding pixel value of the first luminance component image data; creating corrected edge image data with each difference value of the edge image data corrected under given conditions; creating a third luminance component-image data by adding each offset value of the corrected edge image data to its corresponding pixel value of the second luminance component image data; and applying a third filtering step to the color-difference component image data with each pixel of the color-difference component image data being designated as a target pixel, thereby creating a first color-difference component image data with its entirety smoothened.

0011With the above image noise reducing method, the step of creating the first luminance component image data is a process for reducing block noise due to luminance difference, and the step of creating the third luminance component image data is a process for reducing mosquito noise due to color difference. Also, the step of creating the first color-difference component image data is a process for reducing both block noise due to color difference and mosquito noise due to color difference.

0012The block noise reducing process is first performed because if the mosquito noise reducing process, which is a kind of smoothing process, is first performed, block noise is entirely smoothened and it is hard to reduce only block noise in the subsequent block noise reducing process.

0013In the mosquito noise reducing process, not only the second luminance component image data is created, but also the third luminance component image data is created by combining the corrected edge image data created from the edge image data with the second luminance component image data. This is because only a small luminance difference can be smoothened while not greatly smoothing a contour of a large luminance difference, that is, mosquito noise can be reduced without deterioration of the image quality.

0014Then, the third luminance component image data and the first color-difference component image data are finally outputted. These may be used for reproduction or stored in a media, or those with the RGB colors converted may be used for reproduction or stored in a media.

0015In the image noise reducing method of the present invention, clipped values may be used in the first filtering step to have absolute values of the differences in pixel value of each pixel of the filtering range relative to the target value kept within a given threshold value. That is, the first filtering step is performed only for the boundaries of the blocks so that excessive smoothing causes an unnatural (discontinuous) blocks with the boundaries thereof blurred and hence unintentionally emphasizes block noise. In a case where an edge of one block is bright while an edge of an adjacent block is dark, the filtering process may cause excessive correction which exceeds original pixel values. The clipped values are used as the pixel value of the nearby pixels in order to avoid these problems.

0016In the image noise reducing method of the present invention, the corrected edge image data may be created by determining a difference from a maximum difference value and a minimum difference value in the edge image data so that where the difference is greater than a threshold value, each difference value of the edge image data is subtracted or added by a given adjusting value so as to have its absolute value decreased to 0 or greater. Where the difference is greater than the given threshold value, it indicates the possibility that a contour having a great luminance difference exists in the image and therefore mosquito noise is highly likely to have emerged. In order to address this, each difference value of the edge image data is subtracted or added by the given adjusting value so as to have its absolute, value (a luminance difference at its point) decreased. All the difference values of the edge image data are designated as objects to be corrected (which means that the regions with no mosquito noise emerged are corrected), for the reason that if both regions which have been corrected and regions which have not been corrected exist in a block, its boundaries are likely to be noticeable. However, of the difference values of the edge image data, those having absolute values being equal to or lower than the given image edge adjusting value are set at "0" in order to prevent excessive correction for them.

0017Furthermore, in creating the corrected edge image data of the image noise reducing method of the present invention, each difference value of the edge image data may be multiplied by a given adjusting value where the difference is equal to or lower than the threshold value. For the difference being equal to or lower than the threshold value, that is a flat block with less contours existing in the image, it is not meant that there is very little possibility that mosquito noise has emerged.

0018In order to address this, each difference value of the edge image data is multiplied by the given adjusting value to entirely reduce the luminance difference. However, the reduction ratio is set to be relatively moderate compared with a case where the difference is greater than the given threshold.

0019In the image noise reducing method of the present invention, the third filtering step is preferably applied to offset data created by creating difference data by calculating the difference between the pixel value of each target pixel and the pixel value of each pixel within the filtering range and clipping the difference data at an upper limit and a lower limit in a given upper-lower-limit table. According to the visual characteristics, the human eye is not sensitive to color difference. Therefore, no specific problem may arise even when the third filtering step is applied to the color-difference component image data itself. However, by first creating the offset data from the color-difference component image data and then applying the third filtering step thereto, occurrence of unnecessary color blurring can be properly prevented.

0020In the image noise reducing method of the present invention, the given upper-lower-limit table may be created so that for an input value having an absolute value lower than a given threshold value, this input value is designated as an output value, and for an input value having an absolute value equal to or greater than the given threshold value, a threshold value of the same code as that of this input value is designated as an output value.

0021As described above, the image noise reducing method of the present invention realizes reducing of both, block noise and mosquito noise. In the block noise reducing process, the image is previously divided into blocks and the boundaries of the blocks are smoothened. Therefore, unlike a conventional process, contours on the background of an image are not erroneously recognized as boundaries and subjected to the block noise reducing process. Thus, block noise is properly reduced without deterioration of the image quality. In the mosquito noise reducing process, unlike a conventional process which is ended up only by realizing the luminance smoothing, the correction is applied to the image data with its luminance smoothened so that mosquito noise can be properly reduced without causing deterioration of the image, quality. Thus, both block noise and mosquito noise can be properly reduced without deterioration of the image quality. Furthermore, the image noise reducing method of the present invention is performed so that the luminance component image data and the color-difference component image data are processed independently of each other. As a result, it is expected to more securely reduce image noise. <ul id="ul0001" list-style="none"><li>FIG. 1 is a structural view of an image processing apparatus according to one embodiment of the present invention.</li><li>FIG. 2 is a flowchart of the image processing of this embodiment.</li><li>FIG. 3 is a flowchart of an image noise reducing process of FIG. 2.</li><li>FIG. 4 is an explanatory view illustrating a state in which boundary lines have been added to an image data.</li><li>FIGS. 5A and 5B are explanatory views illustrating a state in which boundary lines have been added to image data along the horizontal axis. Specifically, FIG. 5A illustrates a case in which the width of the image data is a multiple of the width of a block, and FIG. 5B illustrates a case in which the width of the image data is not a multiple of the width of a block.</li><li>FIGS. 6A and 6B are explanatory views illustrating a state in which boundary lines have been added to image data along the vertical axis. Specifically, FIG. 6A illustrates a case in which the height of the image data is a multiple of the height of a block, and FIG. 6B illustrates a case in which the height of the image data is not a multiple of the height of the block.</li><li>FIGS. 7A and 7B are flowcharts of the block noise reducing process of FIG. 3, in which FIG. 7A is for a luminance component image data and FIG. 7B for a color-difference component image data.</li><li>FIGS. 8A and 8B are explanatory views of the block noise reducing process in the vertical direction, in which FIG. 8A illustrates a state in which a filtering process is to be performed and FIG. 8B illustrates a state in which vertical block noise has been reduced.</li><li>FIGS. 9A and 9B are explanatory views of the block noise reducing process in the horizontal direction, in which FIG. 9A illustrates a state in which a filtering process is to be performed and FIG. 9B illustrates a state in which horizontal block noise has been reduced.</li><li>FIG. 10 is an explanatory view of an upper-lower-limit table of the color difference for use in the block noise reducing process to the color-difference component image data of FIG. 3.</li><li>FIGS. 11A-11E are explanatory views of the block noise reducing process to the color-difference component image data of FIG. 3, in which FIG. 11A:color-difference component image data, FIG. 11B:difference value data, FIG. 11C: offset data, FIG. 11D: offset data and FIG. 11E illustrates a state in which the pixel value of a target pixel has been replaced.</li><li>FIG. 12 is a flowchart of the mosquito noise reducing process of FIG. 3.</li><li>FIG. 13 is an explanatory view of the edge image creation step of FIG. 12.</li></ul>

0022Now, the description will be made for the structure of an image processing apparatus that realizes the method of reducing noise in images according to one embodiment of the present invention with reference to the drawings attached hereto. The image processing apparatus comprises a computer with a CPU 1, a ROM 2, a working memory 3, a frame memory 4, a data input-output unit 5 and a hard disk 6, which are all connected to a bus 7. The ROM 2 serves to store an image noise reducing program, other computer programs and various parameters, while the working memory 3 that is required for realizing control by the CPU contains such as a buffer and register. The CPU 1 porforms various calculations and processes based on computer programs stored in the ROM 2.

0023The frame memory 4 is a memory for storing image data obtained by decoding a still image compressed and encoded in JPEG format. Image data (R, G, B) inputted in the data input-output unit 5 are once stored respectively in separate frame memories 4, as R component image data, G component image data and B component image data, and then the image noise reducing process is performed. Upon the finish of the image noise reducing process, the (R, G, B) image data are outputted to the outside via the data input-output unit 5 or stored in the hard disk 6.

0024As illustrated in FIG. 2, in the image noise reducing process, a RGB/YCC data conversion process (S1) is first performed and then an image noise reducing process for block noise and mosquito noise (S2) is performed. In the RGB/YCC data conversion process, the (R, G, B) image data are color converted into (Y, Cr, Cb) image data based on the following equations (Eq. 1-Eq. 3). The reason why the data are converted into YCC color space is that block noise and mosquito noise are generated when the JPEG format performs compression/expansion in YCC color space, and therefore correction accuracy is improved when the image noise reducing process is performed in the same color space.<maths id="math0001" num="(1)"><math display="block"><mrow><mtext>Y=(RToY[0][0] × R+ RToY[0][1] × G+ RToY[0][2] × B)/10000</mtext></mrow></math><img file="EP1549075A2_D0001.tif" /></maths><maths id="math0002" num="(2)"><math display="block"><mrow><mtext>Cr=(RToY[1][0] × R+ RToY[1][1] × G+ RToY[1][2] × B)/10000+2048</mtext></mrow></math><img file="EP1549075A2_D0002.tif" /></maths><maths id="math0003" num="(3)"><math display="block"><mrow><mtext>Cb=(RToY[2][0] × R+ RToY[2][1] × G+ RToY[2][2] × B)/10000+2048</mtext></mrow></math><img file="EP1549075A2_D0003.tif" /></maths>       RToY[i][j]:YCrCb conversion coefficient

0025Upon the finish of the image noise reducing process (S2), a YCC/RGB data conversion process (S3) is performed based on the following equations (Eq. 4-Eq. 6) to return the (Y, Cr, Cb) image data to the (R, G, B) image data. Thus, a series of the processes are finished.<maths id="math0004" num="(4)"><math display="block"><mrow><mtext>R=(YToR[0][0] × Y+YToR[0][1] × (Cr-2048)+YToR[0][2] ×(Cb-2048)/10000</mtext></mrow></math><img file="EP1549075A2_D0004.tif" /></maths><maths id="math0005" num="(5)"><math display="block"><mrow><mtext>G=(YToR[1][0] × Y+YToR[1][1] × (Cr-2048)+YToR[1][2] × (Cb-2048)/10000</mtext></mrow></math><img file="EP1549075A2_D0005.tif" /></maths><maths id="math0006" num="(6)"><math display="block"><mrow><mtext>B=(YToR[2][0] × Y+YToR[2][1] × (Cr-2048)+YToR[2][2] × (Cb-2048)/10000</mtext></mrow></math><img file="EP1549075A2_D0006.tif" /></maths>       YToR[i][j]:Y coupling coefficient

0026In this embodiment, in order to limit data loss due to the above processing, the density data format is upgraded from 8 bit to 12 bit, although it is possible to keep the density data format in 8 bit.

0027As illustrated in FIG. 3, in the image noise reducing process of S2, a boundary interpolation step (S20) is performed, and then a block noise reducing step (S21) and a mosquito noise reducing step (S22) are subsequently performed. The boundary interpolation step (S20) is performed for the purpose of interpolating pixel data on the boundaries of the blocks of the image data at the time of performing a later-described filtering step to three image data sets (Y-image data of (Y, Cr, Cb) image data (hereinafter referred to "luminance component image data"), Cr-image data and Cb-unage data (hereinafter respectively referred to "color-difference component image data")).

0028Specifically, as illustrated in FIG. 4, which illustrates any one of the image data sets (each grid and a numerical value in each grid respectively representative of a pixel and a pixel value), the boundary interpolation step (S20) adds to boundaries S of blocks B to be encoded and decoded (non-hatched blocks defined with thick frame) of image data A (a non-hatched portion), blocks B' (hatched blocks with thick frame) each having the same size as the size of each block B respectively in the vertical and horizontal directions, and embeds the pixel values of the pixels on the boundaries of the image data A in the grids (pixel values) of the blocks B'. These added blocks B' are removed in a boundary block removing step (S23) upon the finish of the block noise reducing step (S21) and the mosquito noise reducing step (S22).

0029As illustrated in FIG. 5A, in a case where the length (number of pixels) W of the image data A along the horizontal axis is a multiple of the width (number of pixels) of the block B along the horizontal axis, or a multiple of 8, the blocks B' are added so as to allow each row to have a overall length (number of pixels) of [W+16] Otherwise (when not a multiple of 8), the blocks B' and a surplus are added so as to allow each row to have a overall length (number of pixels) of [W+(8-(W mod 8)) +16], thus achieving data interpolation to a surplus B" of the multiple of 8, as illustrated FIG. 5B.

0030Similarly, as illustrated in FIG. 6A, in a case where the length (number of pixels) H of the image data A along the vertical axis is a multiple of the length (number of pixels) of the block B along the vertical axis, or a multiple of 8, the blocks B' are added so as to allow each column to have a overall length (number of pixels) of [H+16]. Otherwise (when not a multiple of 8), the blocks B' and a surplus are added so as to allow each column to have a overall length (number of pixels) of [H+(8―(H mod 8)) +16], thus achieving data interpolation to a surplus B" of the multiple of 8, as illustrated FIG. 6B.

0031The block noise reducing step (S21) is to create luminance component image data Y1 for luminance component image data Y0 by performing a vertical block noise reducing step (S30) and then create luminance component image data Y2 (first luminance component image data of the present invention) by performing a horizontal block noise reducing step (S31), as illustrated in FIG. 7A. On the other hand, for each of color-difference component image data Cr0, Cb0, a color-difference upper-lower-limit table creation step (S40) is performed. Then, a horizontal color smoothing step (S41) is performed to create color-difference component image data Cr1, Cb1, and then a vertical color smoothing step (S42) is performed to create color-difference component image data Cr2, Cb2 (first color-difference component image data of the present invention).

The Vertical Block Noise Reducing Step (S30)

0032A one-dimensional filter in a horizontal direction is applied to image data with a pixel on the boundaries of the blocks being designated as a target pixel, thereby creating the luminance component image data Y1 that has a luminance difference in the horizontal direction of luminance component image data Y0 eliminated or smoothened in the boundaries of the blocks. The filter size can be set to such as 3 pixels or 5 pixels (see FIG. 8A, in which a dense dot pattern C, a thin dot pattern D and a thinner dot pattern E respectively represent a target pixel, a filtering range and pixels to be filtered), and a filtering step (a first filtering step of the present invention) is performed by using the following equation (Eq. 7).<maths id="math0007" num="(7)"><math display="block"><mrow><mtext>Y1= </mtext><mfrac><mrow><mtext>ΣFb × Yi </mtext></mrow><mrow><mtext>ΣFb</mtext></mrow></mfrac></mrow></math><img file="EP1549075A2_D0007.tif" /></maths>

0033This filter is a weighting filter by matrix Fb (e.g., Fb=(1 <u>3</u> 1)) in which a coefficient to a target value is about 10-20 times greater than a coefficient to the other pixels, and is made based on an equation in which coefficients of the matrix Fb are respectively multiplied by pixel values (Yi) within the filtering range with the target pixel C designated as the center and the results are summed up, and then the sum is divided by the sum of the coefficients of the matrix Fb.

0034The above filtering process, targets of which are only the boundaries between the blocks, may cause an unnatural (discontinuous) block with the boundaries thereof blurred and hence unintentionally emphasize block noise when smoothing is excessively made, and may cause excessive correction which exceeds original pixel values in a case where an edge of one block is bright while an edge of an adjacent block is dark. In order to avoid this problem, in the above equation (Eq. 7), clipped (gap-processed) values are used to meet the requirements of the following equation (Eq. 8), that is, to have absolute values of the differences in pixel value relative to the target value C kept within a threshold value b.<maths id="math0008" num="(8)"><math display="block"><mrow><mtext>Yi[x,y]-b≦Yi'[x-1,y]≦Yi[x,y]+b</mtext></mrow></math><img file="EP1549075A2_D0008.tif" /></maths>

0035The threshold value b is for example 30 so that the pixel values of pixels E on the boundaries of the blocks are converted from the values of FIG. 8A into the values of FIG. 8B, from which it has been found that the difference in pixel value (luminance difference) in the boundaries of the blocks became smaller.

The Horizontal Block Noise Reducing Step (S31)

0036A one-dimensional filter in a vertical direction is applied to image data with a pixel on the boundaries of the blocks being designated as a target pixel, thereby creating the luminance component image data Y2 that has a luminance difference in the vertical direction of the luminance component image data Y1 eliminated or smoothened in the boundary of the blocks. The processing is substantially the same as in the vertical block noise reducing step.

0037The threshold value b is for example 30 so that the pixel values of pixels E on the boundaries of the blocks are converted from the values of FIG. 9A into the values of FIG. 9B, from which it has been found that the difference in pixel value (luminance difference) in the boundaries of the blocks became smaller.

0038Thus, the above two block noise reducing steps are performed for the purpose of reducing the luminance difference in the boundaries of the blocks both in the vertical and horizontal directions by finally creating the luminance component image data Y2 (FIG. 9B) from the luminance component image data Y0 (FIG. 8A). According to the visual characteristics, the human eye is very sensitive to the luminance difference and therefore block noise results mainly from the luminance difference among pixels on the boundaries of the blocks. In light of this, the above two block noise reducing steps are very effective processes that can reduce block noise. However, block noise results not only from the luminance difference but also from the color difference to some extent. In order to completely reduce block noise resulting from these differences, the following color smoothing steps (S41, S42) will be needed.

The Color-Difference Upper-Lower-Limit Table Creation Step (S40)

0039In the color smoothing steps, a relatively large filter (hereinafter described) is used so that where excessive smoothing is made, colors are blurred. For example, where color smoothing is made in great span extending for example from a red of the lips to the skin of a person in image data, the colors are blurred. In order to avoid this problem, the color-difference upper-lower-limit table is used in the color smoothing steps (S41, S42).

0040The color-difference upper-lower-limit table is a table for calculation of mask values (see FIG. 10). A reference code "c" represents a threshold value for the upper and lower limits of an output color difference, and lies in the range of 0-4095 According to this table, for input values falling in the range of -c to c, the corresponding or equivalent values are outputted, while for input values lower than -c, all are set to -c and then this -c is outputted.

The Horizontal Color Smoothing Step (S41)

0041A one-dimensional filter in a horizontal direction is applied to image data with each pixel of a block being designated as a target pixel, thereby creating color-difference component image data Cr1, Cb1, each having colors of color-difference component image data Cr0, Cb0 smoothened in the horizontal direction. The filter size can be set to such as 7 pixels (see FIG. 11A, in which a thick frame C represents a target pixel).

0042Upon calculation of the difference (-2047 to 2047) between the target pixel C and a nearby pixel, difference data (A) is created (FIG. 11B). Then, offset data (B) is created by having this difference data (A) clipped at an upper limit and a lower limit in the upper-lower-limit table (a threshold value c: e.g., 15) (FIG. 11C). Then, upon creation of offset data (C) by returning the pixel value of a target pixel C to the target pixel C of this offset data (B) (FIG. 11D), a filtering step (a third filtering step of the present invention) is performed by using the following equation (Eq. 9) (FIG. 11E).<maths id="math0009" num="(9)"><math display="block"><mrow><mtext>Cr1, Cb1 = </mtext><mfrac><mrow><mtext>ΣFc × Ci </mtext></mrow><mrow><mtext>ΣFc</mtext></mrow></mfrac></mrow></math><img file="EP1549075A2_D0009.tif" /></maths>

0043This filter is moving average filter by matrix Fc (Fc=(1 1 ... 1)) having an equal coefficient, and is made based on an equation in which the coefficient of the matrix Fc is multiplied by respective offset values (Ci) with the target pixel C designated as the center and the results are summed up, and the sum is divided by the sum of the coefficients of the matrix Fc (this moving average filter necessitates the sum of the coefficients to be equal to the filter size). This filtering step is performed for every pixel.

The Vertical Color Smoothing Step (S42)

0044A one-dimensional filter in a vertical direction is applied to image data with each pixel of a block being designated as a target pixel, thereby creating color-difference component image data Cr2, Cb2, each having colors of color-difference component image data Cr1, Cb1 smoothened in the vertical direction. The processing is substantially the same as in the horizontal color smoothing step.

0045Thus, the above two color smoothing steps are performed for the purpose of smoothing or losing only small color differences while leaving a contour having a large color difference unsmoothened by finally creating the color-difference component image data Cr2, Cb2 from the color-difference component image data Cr0, Cb0 (FIG. 11A). That is, where no upper and lower limits are provided for the magnitude of the color difference, smoothing is made based on a normal moving average, thereby causing a contour having a large color difference to be blurred. Instead, where a threshold value c is set so as to have upper and lower limits lying in the fluctuation range of the magnitude of a small color difference, the magnitude of a color difference resulting from a contour having a large color difference located in the periphery is transformed to lie in the fluctuation range of the magnitude of a small color difference of a contour so as to prevent the contour having a large color difference from being blurred. Block noise due to color difference is caused by this small color difference and therefore the above two color smoothing steps are very elective processes that can reduce block noise due to color difference with no deterioration of the image quality or with making deteriorated image quality non-noticeable. The adjustment of intensity is made by properly setting the filter size of each filter and the threshold value of the upper-lower-limit table.

0046As illustrated in FIG. 12, the mosquito noise reducing step (S22) is to create luminance component image data Y3 (second luminance component image data of the present invention) by performing a luminance smoothing step (S50) for the luminance component image data Y2 created in the block noise reducing step of S21, then create edge image data E1 by performing an edge image data creation step (S51) based on the luminance component image data Y2, then create corrected edge image data E2 by performing edge image data correction step (S52) and then finally create luminance component image data Y4 (third luminance component image data of the present invention) by combining this corrected edge image data E2 with the luminance component image data Y3 in a step (S53).

The Luminance Smoothing Step (S50)

0047A two-dimensional filter is applied to image data with each pixel of a block being designated as a target pixel, thereby creating the luminance component image data Y3 that has a luminance of the luminance component image data Y2 smoothened. The filter size can be set to such as 3 by 3 pixels or 5 by 5 pixels. A filtering step (a second filtering step of the present invention) is performed by using the following equation (Eq. 10).<maths id="math0010" num="(10)"><math display="block"><mrow><mtext>Y3=Y2+( </mtext><mfrac><mrow><mtext>ΣFb × Yi </mtext></mrow><mrow><mtext>ΣFb</mtext></mrow></mfrac><mtext>- Y2) × (d/128)</mtext></mrow></math><img file="EP1549075A2_D0010.tif" /></maths>

0048This filter is a moving average filter by matrix Fb having an equal coefficient, and is made based on an equation in which the coefficient of the matrix Fb is multiplied by respective pixel values (Yi) within the filtering range and the results are summed up; the sum is divided by the sum of the coefficients of the matrix Fb; the difference between this calculated value and the pixel value of the target pixel C is determined; this determined value is then multiplied by [d/128]; and this calculated value is added to the pixel value of the target pixel C. The "d" represents a coefficient (smoothing intensity coefficient) for adjusting the smoothing intensity. The smoothing intensity coefficient d is divided by 128 because the smoothing intensity coefficient d is previously multiplied by [128/100] for high speed processing.

The Edge Image Data Creation Step (S51)

0049The edge image data E (see FIG. 13) is created for each block B by using the following equation (Eq, 11), in which the data is clipped to the range of -2048 to 2047.<maths id="math0011" num="(11)"><math display="block"><mrow><mtext>E1=Y2-Y3</mtext></mrow></math><img file="EP1549075A2_D0011.tif" /></maths>

0050That is, the edge image data E1 is created by subtracting each pixel value of the luminance component image data Y3 created in the luminance smoothing step (S50) from a corresponding pixel value of the luminance component image data Y2.

The Edge Image Data Correction Step (S52)

0051A difference SA is determined from the maximum difference value and the minimum difference value in the edge image data E1 and corrected edge image data E2 is created by using the following equation (Eq. 12).<maths id="math0012" num=""><img file="EP1549075A2_D0012.tif" /></maths>

0052Where the difference SA is greater than a threshold value e, it indicates the possibility that a contour having a great luminance difference exists in the image and therefore mosquito noise is highly likely to have emerged. In order to address this case based on the above equation (Eq. 12), the following calculation is made. That is, mosquito noise emerges in a region having an excessively large luminance difference and therefore each difference value of the edge image data E1 is subtracted or added by an image edge adjusting value "f" of mosquito noise so as to have its absolute value (a luminance difference at its point) decreased. All the difference values of the edge image data E1. are designated as objects to be corrected (which means that the regions with no mosquito noise emerged are also corrected), for the reason that if both regions which have been corrected and regions which have not been corrected exist in a block, its boundaries are likely to be noticeable. However, of the difference values of the edge image data E1, those having absolute values being equal to or lower than the, image edge adjusting value f are set at "0" in order to prevent excessive correction for them.

0053For the difference SA being equal to or lower than the threshold value, that is, a flat block with less contours existing in the image, it is not meant that there is very little possibility that mosquito noise has emerged. In order to address this case based on the above equation (Eq. 12), the following calculation is made. That is, all the individual difference values of the edge image data E1 are multiplied by [1/an image edge adjusting value "g" of a normal image] to entirely reduce the luminance difference. However, the reduction ratio is set to be relatively moderate compared with a case where the difference SA is greater than the threshold value.

0054The threshold value e is for example 10, while the image edge adjusting value f of mosquito noise and the image edge adjusting value g of the normal image g are each for example 5. With these values applied, in a case of FIG. 13, the difference SA is 179 so that a target pixel (difference value: -4) at the upper left corner is: 4 -5=-1→0 after the correction, its adjacent target pixel (difference value: -28) is: -28+5=-23, ... a target, pixel (difference value: 7) at the lower right corner is: 7-5=2 after the correction. Thus, the corrected edge image E2 with the luminance differences entirely reduced is created.

The Combining Step (S53)

0055The luminance, component image data Y4 is created by performing a combining step by using the following equation (Eq. 13), in which the data is clipped to the range of 0 to 4096.<maths id="math0013" num="(13)"><math display="block"><mrow><mtext>Y4=Y3+E2</mtext></mrow></math><img file="EP1549075A2_D0013.tif" /></maths>

0056That is, the luminance component image data Y4 is created by adding each offset value of the corrected edge image data E2 created in the edge image data correction step of S52 to its corresponding pixel value of the luminance component image data Y3.

0057The above mosquito noise reducing process is to finally create the luminance component image data Y4 from the luminance component image data Y2, thus smoothing or losing only small luminance differences while leaving contours having a large luminance difference unsmoothened (this is because the luminance difference is originally large and therefore smoothing is not noticeable). Mosquito noise is caused by this small luminance difference and therefore the above mosquito noise reducing process is effective in reducing mosquito noise with no deterioration of the image quality or with making deteriorated image quality non-noticeable.

0058The image noise reducing process of this embodiment is made so that the block noise reducing step (S21) is first performed and then the mosquito noise reducing step (S22) is subsequently performed. This is because if the mosquito noise reducing step (S22), which is a smoothing process, is first performed, block noise is entirely smoothened, and it is hard to reduce only block noise in the block noise reducing step (S21). Therefore, the mosquito noise reducing step is necessarily performed after the block noise reducing step (S21).

0059It is not necessary to limit the present invention to the above embodiment, while various modifications may be made within the scope of the present invention.

0060For example, the color smoothing steps of S41, S42, which are to be performed for the block noise reducing step of S21 in the above embodiment, are effective in smoothing or losing a region of a small color difference, as described above. That is, these steps are also elective for mosquito noise. Accordingly, these color smoothing steps may be made also for the mosquito noise reducing step of S22, while being made along with the color-difference upper-lower-limit table creation step of S40. In either case, by applying the steps of S40-S42 to the color-difference component image data of the image data, image noises in the color-difference component image data are reduced and therefore there is no particular sense in ordering the steps of S40-S42 in the process.

0061The color smoothing steps of S41, S42 of the above embodiment employ the one-dimensional filter since it has a large filter size that realizes a shortened process time. However, where it is not necessary to take into account the process time or the filter size, a two-dimensional filter may be employed. It is not necessary to limit a filter used in the block noise reducing steps of S30, S31 to the one-dimensional filter. For example, it is possible to use a two-dimensional filter of 3 by 3 pixels.

0062In the above embodiment, the block noise reducing steps of S30, S31 employ the weighting filter while the other steps employ the moving average filter. It is not necessary to limit the present invention to this.

0063In the above embodiment, the image nose reducing process is performed on the computer, while it may be incorporated into a coder, allowing the coder to perform the image noise reducing process in the decoding process (more specifically after an inverse orthogonal transform has been performed).

0064In the above embodiment, since the RGB conversion is made at the time of decoding a compressed image, the RGB/YCC data conversion (S1) is performed. For a BMP image or the like whose input image contains RGB information, this RGB/YCC data conversion is required. However, for a JPEG image which originally contains YCC information, the RGB/YCC data conversion (S1), and the YCC/RGB data conversion. (S3) to be performed therealong are not necessarily required.