JP4728411B2 - Method for reducing color bleed artifacts in digital images - Google Patents

Description

  The present invention relates to digital images. More particularly, the present invention relates to decompression of compressed images.

  Image compression with JPEG and some other image compression algorithms converts the image into luminance and chrominance channels, and the chrominance channels (Cb and Cr) are a fraction of an integer (typically 2 or 4). ) Done by sub-sampling. The luminance channel and the subsampled chrominance channel are then compressed by a compression module suitable for grayscale images. The human visual system (HVS) can subsample the chrominance channel because the spatial sensitivity of the color sensation is reduced to about ½ compared to the luminance sensitivity. Thus, in most color photographs that include subsampled chrominance values, the color bleed artifacts resulting from subsampling are substantially inconspicuous.

  However, color bleeding artifacts are noticeable in color compound document images that frequently contain sharp edges between different colors. For example, in a color compound document, text or other computer-generated shapes may overlap on an area having a uniform color. Steep edges occur at the turn of computer-generated images and photographs.

  If the recovered chrominance channels are interpolated back to their original resolution by a simple technique such as linear interpolation, the color bleed artifact is noticeable. Color bleed artifacts are found at the boundary between a region having a pure color and a white region (or a region having a light color). In document images, these boundaries are often between a white or light background and color letters, lines and patches.

  Also, color blurring artifacts are prominent even in areas containing inverted text (for example, white characters in a colored background). Due to the nature of the human visual system, color blurring artifacts appear on the brighter side of the edge.

  Color blur artifacts reduce image quality. Therefore, it is desirable to reduce color bleeding artifacts, especially in color compound document images.

  According to one aspect of the invention, color bleed artifacts in a digital image are reduced by changing the chrominance values of at least some pixels of the digital image. The pixel is changed according to the luminance value and the color dynamic range. Other aspects and advantages of the present invention will become apparent from the following detailed description, taken in conjunction with the accompanying drawings, illustrating by way of example the principles of the invention.

FIG. 6 is a diagram of a method for restoring an image and reducing color blurring artifacts in the restored image. It is a figure of the local pixel vicinity of a pixel. FIG. 4 is a sliding window that determines the color dynamic range of a pixel in a reconstructed image interpolated by pixel replication. FIG. 4 is a sliding window that determines the color dynamic range of a pixel in a reconstructed image interpolated by pixel replication. FIG. 4 is a sliding window that determines the color dynamic range of a pixel in a reconstructed image interpolated by pixel replication. FIG. 4 is a sliding window that determines the color dynamic range of a pixel in a reconstructed image interpolated by pixel replication. FIG. 2 is a diagram of a hardware implementation of the method of FIG.

  As shown in the drawings, the present invention is embodied in a method and apparatus for reducing color blurring artifacts in a digital image that includes subsampled chrominance values. The method and apparatus are particularly suitable for reducing color bleed artifacts in color compound document images that contain both natural features and computer-generated features. Natural features in photographic areas tend to have a mild luminance transition or chrominance transition. Computer-generated features such as text regions tend to have steep edges and steep luminance or chrominance transitions. The method and apparatus reduce color bleed artifacts that occur at sharp edges.

  FIG. 1 illustrates a method for decompressing a compressed image and reducing color blurring artifacts in the decompressed image. The compressed image includes a first stream of compressed luminance (Y) values and second and third streams of compressed chrominance (Cb and Cr) values. Prior to compression, the chrominance channel is subsampled or low-pass filtered and then subsampled.

  Image decompression begins by accessing the first, second and third streams of compressed values (block 102). The luminance value (Y) is restored from the first stream (block 104), and the chrominance values (Cb and Cr) are restored from the second and third streams (block 106).

  Hereinafter, luminance values and chrominance values will be described with respect to normalized values. The luminance value is 0 ≦ Y ≦ 1. The chrominance values are −0.5 ≦ Cb ≦ 0.5 and −0.5 ≦ Cr ≦ 0.5. Point Cb = Cr = 0 corresponds to the gray axis (colorless).

  Since the chrominance values (Cb and Cr) are subsampled during compression, the recovered chrominance channels are matched to the resolution of the luminance channel by interpolating or up-sampling them to their original resolution (block 108). A simple interpolation method is pixel replication, which replaces all low resolution pixels with 2 × 2 blocks of high resolution pixels having the same chrominance values (Cb and Cr). More sophisticated upsampling methods apply low pass filtering to the pixel replicated chrominance channel in order to reduce aliasing artifacts at high spatial frequency locations such as edges.

  Color bleed artifacts in the restored digital image are reduced by changing the chroma values of at least some pixels in the digital image. The pixel's chroma value is changed corresponding to the pixel's luminance value and the local color dynamic range (block 110), and the chrominance values (Cb and Cr) are scaled according to the change in the chroma value (block 112). By changing the chroma value corresponding to the color dynamic range, the chromaticity is not excessively reduced in a region having an essentially uniform light color.

  The restored luminance and chrominance channels can be transformed back to the original resolution RGB space. Then, the restored image can be displayed (for example, printed).

  A pixel's local color dynamic range is determined with respect to its local pixel neighborhood. FIG. 2 provides an example of pixel neighborhoods. The neighborhood outlined by the dashed window contains a 3 × 3 array of pixels. The predetermined pixel marked with “X” is in the center of the neighborhood.

として計算することができ、クロマ値（Ｃ）を

によって概算することができる。この場合、クロミナンス値の各々の許容値範囲は、［−０．５，０．５］である。近傍の色ダイナミックレンジ（Ｄ）は、Ｃ_ｍが近傍の最小クロマ値を表しＣ_Ｍが近傍の最大クロマ値を表す場合、Ｄ＝Ｃ_Ｍ−Ｃ_ｍとして計算してよい。 The color dynamic range of each given pixel can be calculated as the difference between the pixel's minimum and maximum chroma values in the vicinity of the local pixel. The chroma value (C) of the pixel

The chroma value (C) can be calculated as

Can be approximated by In this case, the allowable range of each chrominance value is [−0.5, 0.5]. Near the color dynamic range (D), if C _M C _m represents the minimum chroma values of the neighboring represents the maximum chroma value in the neighborhood may be calculated as D = C _M -C _m.

  The neighborhood and color dynamic range are established for each of the predetermined pixels. Therefore, the color dynamic range is determined in units of pixels.

A new chroma value (C ′) for a given pixel may be calculated as C ′ = C−f (Y, D) · (C−C ₀ ). Where C is the original chroma value of the pixel, D is the color dynamic range, C ₀ is the color coefficient having values from 0 to the minimum chroma near the local pixel of that pixel, and f ( Y, D) is a parameter formula that determines the relative chroma reduction magnitude.

The parameter formula f (Y, D) can vary in the range between 0 and 1. The parameter formula f (Y, D) is subject to the following constraints. That is, if (1) Y → 1, then f (Y, D) → 1, thereby C ′ → C ₀ , and (2) if D → 0, then f (Y, D) → 0. It is. In the case of D → 0, C ′ → C, so that the chroma value of the pixel is not changed.

  The parameter formula f (Y, D) may be subject to another restriction. That is, in the case of D → 0 and Y → 1, f (Y, D) → 0. Thus, the color dynamic range has priority.

  When the chroma value is changed, each corresponding chrominance value (Cb and Cr) is scaled by the ratio C '/ C. Therefore, Cb ′ = C ′ / C · Cb and Cr ′ = C ′ / C · Cr. Thereby, the hue of the color does not change.

Consider different cases for the color coefficient (C ₀ ). In the first case, C ₀ = C _m . At this time, C ′ = C−f (Y, D) · (C−C _m ), and the new chroma value (C ′) is the original chroma value (C) and the minimum neighborhood value (C _m ). between. In the case of high luminance Y → 1, the new chroma value (C ′) approaches the local minimum value (C _m ). Therefore, the C '→ _{C m} accordance Y → 1.

In the second case, C ₀ = max (C _m −D, 0). With such a coefficient C ₀ , the new (C ′) can fall below the local minimum (C _m ). However, the chroma value is still determined by the color dynamic range (D). The color coefficient (C ₀ ) does not decrease below C _m -D. When C _m <D, the color coefficient (C ₀ ) is 0. If the color dynamic range is not small, the situation C _m <D corresponds to the edge between the “no color” region and the “pure color” region. In this situation, the chroma value can be reduced to 0 when the “uncolored” side is close to white. On the other hand, the situation C _m > D may correspond to an edge between “mild-color” and “pure color”, in which case the color blur artifact is not very visible. . In this situation, the chroma value can be reduced to 0, so the color is greatly reduced on the brighter side of the edge. By setting the lower limit C _m -D for the changed chroma value, the chroma value is obtained when the luminance is Y → 1 in the color gradient region (“mild edge”) where the color dynamic range (D) is not so small or large. Is guaranteed not to be significantly reduced.

In the third case, C ₀ = 0. For high brightness Y → 1, the new chroma value (C ′) approaches zero. Therefore, C ′ → 0 according to Y → 1.

  Parameter expressions can be expressed in different ways. A typical parameter formula is f (Y, D) = max [1-α {(1-Y) / D}, 0]. A typical parameter equation always obeys two constraints. That is, f (Y → 1, D) → 1 and f (Y, D → 0) → 0 regardless of the value of the term α (α is positive).

For a given value of D, there is a luminance value (Y) below which the color bleed artifact is not visible. If no color bleed artifact is visible for a given D, the value of the term α can be selected such that f (Y, D) = 0. For example, if Y <(1-D / α), the chroma value is not changed. That is, if the threshold value T _Y (D) = (1−D / α), C ′ = C when Y <T _Y (D).

  The term α may be a function of the color dynamic range (D). The term α may be measured for each value of D. Each term α may be measured from a subjective test (eg, observing the visibility of artifacts in the patch).

  On the other hand, the term α may be constant for all D. The value of the term α can be determined by trial and error. For example, it was found that 0.5 ≦ α ≦ 1 is effective.

  Thus, a method for reducing the amount of chrominance in a pixel resulting in a visual perception of color blur has been clarified. The method does not change patches with high luminance values and non-negligible chrominance values by considering luminance values and local chrominance values. A flat region has a relatively high brightness, but may also have chrominance that cannot be ignored at the same time. The method does not map such areas to white. The document image may have a text box with a light background. The method does not map a light background to white.

  This method is particularly effective in reducing color bleed artifacts in complex color documents. However, the method is not so limited.

  Although FIG. 2 shows a 3 × 3 square neighborhood, the method is not limited to such a neighborhood. The neighborhood is not limited to any particular size. The number of pixels is not limited to nine. It is preferred that the number of neighboring pixels be fixed for all given pixels, but the neighborhood size can be adjusted to accommodate a particular type of image area (eg, text, graphics, natural features). May be changed.

  The neighborhood is not limited to any particular shape, but a square window is preferred for performance reasons. For example, the nearby shape may be a diamond shape. The minimum / maximum calculation may include a central pixel (ie, a perfect neighborhood may be used) or a predetermined central pixel may be excluded from the minimum / maximum calculation (ie, hollow Can be used).

  The method has been described with respect to decomposing a digital image into visually important channels (eg, luminance channels and blue and red chrominance channels). However, the method is not limited to the YCbCr color space. For example, the method can be applied to Yuv, Luv or Lab color spaces (L represents lightness).

  In the method described above, a large amount of calculation is applied to the calculation of the color dynamic range for each neighboring pixel. However, using subsampled chrominance values and thereby utilizing redundancy can reduce the amount of computation and increase the speed of the method.

  Reference is now made to FIGS. 3A-3D, which show sliding windows that determine the color dynamic range of the pixels in the reconstructed image interpolated by pixel duplication. During interpolation, the first pixel of the restored luminance channel is replaced by a 2 × 2 block of pixel C1, and the second pixel of the restored luminance channel is replaced by a 2 × 2 block of pixel C2, and restored. The third pixel of the reconstructed luminance channel is replaced by the 2 × 2 block of pixel C3, and the fourth pixel of the reconstructed luminance channel is replaced by the 2 × 2 block of pixel C4.

The pixels in all 2 × 2 blocks of the upsampled chrominance channel contain exactly the same chrominance value. As shown in FIGS. 3 (a) -3 (d), each 4 × 4 block of an upsampled chrominance channel aligned to a 2 × 2 block boundary has four possibilities to adapt to a 3 × 3 neighborhood. Have sex. For each of these four possibilities, the neighborhood minimum (or maximum) and neighborhood color dynamic range are the same because they are calculated from the same four color coefficient values. The four color coefficient values correspond to four double sized pixels, each pixel corresponding to a subsampled chrominance channel pixel. Thus, computational complexity results in finding the maximum and minimum values near 2 × 2 in the subsampled chrominance channel. This reduces the number of chrominance modulus calculations to ¼, and reduces the number of comparisons required to calculate the neighborhood minimum and maximum chroma values (C _m and C _M ) to 1/16. Is done. When using pixel replication in this way, it is preferable to apply low pass filtering to the chrominance channels prior to their subsampling. For example, all 2 × 2 blocks of the original chrominance channel can be replaced by the average or median of the four chrominance values of that block.

The computational complexity can be reduced in other ways. For example, if the difference between the new chroma value (C ′) and the original chroma value (C) is small, little benefit is gained by changing the chroma value. Therefore, if the difference is small, the parameter formula is not calculated and the chroma value is not changed. For example, assuming that T is a predetermined threshold, if (C−C ₀ ) <T, C ′ = C. The color dynamic range (D) is calculated for all pixels when calculated with a fast recursive algorithm. Therefore, performing the comparison D <T results in little overhead. An exemplary threshold is T = 0.1 (relative to the normalized chrominance value).

  The method can be implemented in hardware, software or a combination of the two.

  Reference is now made to FIG. 4, which shows a software implementation of the method. The computer memory 212 stores a program 214 that instructs the processor 216 to execute the above-described method at runtime. Memory 212 and processor 216 may be part of a personal computer or workstation.

  A non-recursive algorithm can be used to determine the minimum and maximum chroma values near the pixel. Non-recursive algorithms can be used for all types of neighbors. The size of memory required to determine the minimum and maximum values is equal to the neighborhood size.

  A fast recursive algorithm may be used to determine minimum and maximum chroma values for a perfect square neighborhood (eg, 2 × 2, 3 × 3, 5 × 5 neighborhood). Except for the 2 × 2 neighborhood, only 12 comparisons are required to determine the neighborhood's maximum and minimum chroma values, regardless of the neighborhood size. For a 2 × 2 neighborhood, only three comparisons are necessary. When this recursive algorithm is used, the size of memory required to buffer the line is equal to twice the size of the horizontal kernel window.

  The present invention is not limited to the specific embodiments described and illustrated above. Instead, the present invention is construed according to the claims that follow.

Claims (8)

A method for reducing color bleed artifacts in a digital image, the method comprising:
Reducing the chrominance values of at least some pixels in the digital image;
C ₀ is a value between 0 and the minimum chroma value C _m of the local neighborhood of the pixel, C is the original chroma value of the pixel, Y is the luminance of the pixel, D is the minimum value C _{m of the} local neighborhood pixel, and when the color dynamic range is a function of the maximum value C _M, the chrominance value of the number of pixels by subtracting the original chroma value of the pixel by f (Y, D) · ( C-C 0) determined according to obtain chroma, f (Y, D) is a value in the range of 0 and 1, and, Y → 1 when f (Y, D) a → 1, when D → 0 f Said method wherein (Y, D) → 0.   The method of claim 1, wherein the f (Y, D) also follows f (Y, D) → 0 if D → 0 and Y → 1. The method of claim 1, wherein the color dynamic range D = (C _M −C _m ).   If the original chroma value (C) of a pixel is changed and C ′ is the new chroma value, the original chrominance value of the pixel is scaled by the ratio C ′ / C to obtain a new chrominance value. The method of claim 1, wherein: An apparatus for reducing color bleed artifacts in a digital image, the apparatus comprising means for reducing chrominance values of at least some pixels in the digital image;
The means for reducing the chrominance value comprises: C ₀ is a value between 0 and a minimum chroma value C _{m in} the local vicinity of the pixel, C is the original chroma value of the pixel, Y is the luminance of the pixel, and D is a local value. When the color dynamic range is a function of the minimum value C _m and the maximum value C _M of the neighboring pixels, the chrominance values of the several pixels are set as f (Y, D) · ( C (C ₀ ) is determined according to the chroma value obtained by subtraction, and f (Y, D) is a value in the range of 0 and 1, and f (Y, D) → 1 when Y → 1 , D → 0, where f (Y, D) → 0.   6. The apparatus of claim 5, wherein the f (Y, D) also follows f (Y, D) → 0 if D → 0 and Y → 1. The apparatus of claim 5, wherein the color dynamic range D = (C _M −C _m ).   The means for reducing the chrominance value is such that if the original chroma value (C) of a pixel is changed and C ′ is a new chroma value, the original chrominance value of the pixel is expressed as a ratio C ′ / C. The apparatus of claim 5, wherein the apparatus is configured to scale to determine a new chrominance value.

Images