JP2005198219A - Estimation of frequency and amplitude of halftone screen for digital screening cancel of document - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an efficient method and system for estimating a screen frequency and amplitude. <P>SOLUTION: An efficient method and system for eliminating halftone screens from scanned documents while preserving the quality and sharpness of text and line-art is disclosed. The method and system utilizes one or more independent channels with different sensitivities (e.g., Max, High, and Low) to provide high quality frequency and amplitude estimation. The most sensitive channel (Max) derives the frequency estimate, and the remaining channels (e.g., High and Low) are combined to create the screen amplitude. The Max channel is the most sensitive and will usually report the existence of frequencies even when the screen is very weak. Therefore, the screen frequency must be additionally qualified by the screen amplitude. The screen amplitude can be interpreted as the level of confidence that the local neighborhood represents half-toned data. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a method for obtaining an estimation of a screen frequency and a magnitude of an image signal, or an estimation device for a screen frequency and a magnitude.

  Almost all prints except silver halide photography are printed using a halftone screen. The need to estimate halftone frequency and size is due to the fact that almost all prints except some devices such as sublimation or silver halide photography are printed out using a halftone screen. Arise. These halftones are very unique to the printing device and will result in visible artifacts and / or unacceptable moire patterns if not properly removed when scanned and halftoned again. Halftones are typically printed with four or more color separations with slightly different screens at different angles and / or frequencies, which interfere with each other and create undesirable spatial artifacts. Suppressing halftones is particularly important for color documents as it will result.

  Successful removal of the original halftone screen is based on the ability to accurately estimate the local frequency. Accordingly, there is a need for an improved method and apparatus for estimating the frequency and size of a halftone screen.

  An effective method and system for removing a halftone screen from a scanned document while preserving the quality and sharpness of text and line art is disclosed. This method and system uses one or more independent channels with different sensitivities (eg, Max, High and Low) to give a high quality estimate of frequency and magnitude. The most sensitive channel (Max) derives an estimate of the frequency and the remaining channels (eg, High and Low) are combined to give the screen size. The largest channel is the most sensitive and will usually signal the presence of a frequency even when the screen is very weak. Therefore, the screen frequency must be additionally limited by the size of the screen. The size of the screen can be interpreted as a level of confidence that the local neighborhood represents halftoned data.

  A new method and system for unscreening digitally scanned documents to eliminate or substantially reduce potential halftone interference and unwanted moiré patterns is described. Referring to FIG. 1, a block diagram of the method and system of the present invention is represented by a screen estimator module SEM. The screen estimator module is responsive to estimate the instantaneous halftone frequency and magnitude (intensity) at the current pixel of interest. The screen estimator module operates on the 8-bit source image Src28 and generates an 8-bit halftone frequency estimate Scf70 and an 8-bit halftone magnitude estimate Scm72.

  The need for halftone frequency and size estimation is the fact that almost all prints, except some devices such as sublimation or silver halide photography, are printed out using a halftone screen. Arise from. These halftones are very specific to the printing device and can cause visible artifacts and / or unacceptable moire patterns if not properly removed when scanned for printing and halftoned again. Will occur. A screen releaser module (DSC) as described in Applicant's pending application, Applicant Serial Number D / A 3010, removes (filters) the original halftone pattern from the original scanned image. ) To rely on the information generated by the screen estimator module. Halftones are typically printed with four or more color separations with slightly different screens at different angles and / or frequencies, which interfere with each other and are undesirable spatial artifacts. Suppressing halftones is particularly important for color documents.

  The screen estimator module consists of one or more frequency estimates (eg, Mx, Hi and Lo) acting in parallel. The first Mx channel 30 is used to estimate the screen frequency Scf70. The remaining channels (e.g., Hi40 and Lo50, respectively) are combined together at the last minute to generate a screen magnitude signal Scm72. High quality estimation of small point color halftoned text requires the use of two or more channels to estimate the screen size to cover the desired frequency range of interest. is there.

  In one embodiment, the screen estimator module SEM may use up to three frequency channels with different sensitivity levels. The upper Mx channel 30 of FIG. 1 is adjusted for maximum frequency sensitivity at full source resolution, and is thereby used to derive the screen frequency estimation signal Scf70. However, this channel is very sensitive and will usually report the presence of a frequency even when the screen is very weak. Therefore, the screen frequency must be additionally limited by the screen size Scm72.

  The Hi channel 40 in FIG. 1 is tuned for moderate frequency sensitivity that is less sensitive than the Mx channel 30. The two channels Mx30 and Hi40 operate at full source resolution. The Lo channel 50 is also tuned for moderate sensitivity. However, in contrast to the Mx30 and Hi40 channels, the Lo50 channel operates on the sub-sampling signal output from the triangular Lo channel filter F3 / 256 at half the source resolution in each direction. The screen magnitude signal Scm72 is derived from an analysis of one or more frequency estimates generated by the Hi40 and Lo50 channels.

  Each frequency channel is composed of a plurality of Min-Max texture detectors MM3 31, 32 and 33, which will be described later, followed by average filters 41, 42 and 52. The Mx30 and Hi40 channel MM3 32 units operate on a single channel 8-bit incoming source signal Src28, while the Lo channel MM3 32 operates on a subsampled signal at half resolution. The Lo channel F3 / 2 filter 56 is responsive to filter the source signal Src28, subsample by 2x coefficients in each direction, and activate the Lo channel MM3 unit 32.

  Three MM3Min-Max modules 31, 32 and 33 are used to find peaks and valleys in the 2D input signal. Since the Mx30 channel and the Hi40 channel share the same Src input signal 28, they overlap the first stage calculation of the MM3 unit 32. However, different thresholds are applied in the second stage of the two units, giving two distinct results. The dashed line 27 in FIG. 1 is intended to serve as a reminder that the front end portions of the two MM3 units 32 can be calculated once and then shared.

  A detailed description of the Min-Max detector unit is given below. The unit basically examines the contents of a 3x3 window centered on the current pixel of interest and applies adaptive thresholding if the center pixel is significantly larger or smaller than the eight neighboring pixels that surround it. Use to analyze. In that case, the central pixel is considered to be a peak (if large) or a valley (if small), respectively. By counting the number of peaks and valleys per unit area, a measure of the local frequency is obtained.

  The outputs 31, 32 and 33 of each MM3 unit have a precision of only 1 bit, but each is scaled by the configuration factor DotGain before the first subsequent filtering. Each unit operates one or more color channels of the input signal. However, in this embodiment, only one channel is used, the luminance channel. The DotGain coefficient for Lo channel 50 is divided by some coefficient, such as four. Note that this scaling can be postponed to the normalization step of the first subsequent filter by adjusting the normalization factor at that stage.

  The outputs from the MM3 Min-Max detectors 31, 32 and 33 are passed through different averaging and subsampling filters. In order to prevent problems with subsampling, the spatial filter span in each case is doubled from the subsampling ratio minus one. The Mx channel 30 uses a triangular 2D F63 / 32 filter 32 that reduces the bandwidth by a factor of 32x in each direction (approximately one-thousandth of the source bandwidth).

  Similarly, the output of MM3 of Hi channel 40 is applied in cascade to two triangular 2D subsampling filters, namely F31 / 16 filter 42 and F3 / 2 filter 46. The output from the cascaded filter processing unit is also subsampled by 32x coefficients in each direction (16x for the first filter and 2x for the second filter) so that the output is at the same data rate as the Mx channel 30. is there.

Similarly, Lo channel 50 uses two triangular 2D filters, F15 / 852 and F3 / 246, in a cascaded fashion. The output from the second filtering unit is also subsampled by 32x coefficients in each direction (first 2x, then 8x and 2x by F3 / 2). Is more bandwidth data path of the high-order is shown in Figure 1, thick black lines in the source bandwidth, narrow lines by 1/16 bandwidth, the thin black line 1/32 ^nd bandwidth used . The reduction factor is also specifically indicated by a number.

  In both the Hi30 channel and the Lo40 channel, a 1/16 resolution sample signal is sent to the MX3 unit 44. These perform 3 × 3 Max operations (gray expansion). The output is sent to each b input of the channel duplex bilinear interpolation unit DBI 54, respectively.

  Unlike the Mx channel 30, the Hi40 and Lo50 size estimation channels include an additional smoothing / average F5 64 stage to further reduce spatial noise. The F5 unit 64 is a 5 × 5 triangle weight (non-subsampling) filter. The filtered outputs from these units are sent to the inputs of their respective double bilinear interpolation units DBI 54. The output is also sent through a C3 contrast unit 48 that finds the largest difference in a 3 × 3 window centered on the current pixel. The C3 outputs are respectively c inputs to the DBI unit 48.

  The Mx channel 30 averaged at 1/32 resolution is sent to the bilinear interpolation unit SCF 36. The three signals generated by each of the Hi 40 and Lo 50 channels are sent to their respective DBI units 54. These units perform double bilinear interpolation that brings the subsampled input resolution back to the original source resolution. The DBI a and c inputs are at 1/32 resolution, and the b input is at 1/16 resolution. The output bandwidth from the interpolation unit is substantially higher than the input. For example, for coefficients higher than 32x, the interpolation unit produces 1024 output pixels for each input pixel.

  The interpolated output of the interpolation unit SCF 36 of the Mx channel 30 is an 8-bit estimated screen frequency Scf70. The outputs of other channels, such as Hi40 and Lo50 channel double interpolation units (Hi and Lo), are combined with each other in the magnitude estimation module SCM61. This output is an 8-bit estimated screen size signal Scm72. The estimated screen frequency and magnitude signals Scf70 and Scm72 are exported to the screen cancellation module DSC and (only Scm) are exported to the segmentation module SEG (both not shown). A more detailed description of the various elements of the screen estimation module is given below.

  FIG. 2 shows the one-dimensional filter response of the various filter units, and FIGS. 3-5 show the two-dimensional filter response of the various units. These filtering units are used for the purpose of smoothing or averaging the input signal to remove high frequencies. Each filter unit implements a square separable symmetrical 2D FIR (Finite Impulse Response) filter. The filter response is the same in the horizontal and vertical directions. If the input to the filter is a color signal, the same filter response is independently applied to each one of the color components. The 1D filter 60 response is a symmetric triangle with integer coefficients as shown in FIG. A specific filter shape (but other filter shapes are also covered) was chosen for convenience of implementation.

  A common filter form is called an Fn / k filter, where n is the filter size (overall span in the x or y direction) and k is the sub-sampling amount applied to the output filtered in each direction. is there. The sub-sampling coefficient k is omitted when k = 1. In this document, the filter span n is an odd integer (n = 1, 3, 5...) So that the 2D filter response has a well-defined peak at the effective center pixel location.

  Examples of 1D and 2D filter responses are shown in FIGS. FIG. 2 shows the response of the denormalized 1D filter 60 to F3 and F11, and FIGS. 3 to 5 show the resulting denormalized 2D coefficients for F3 62, F5 64, and F7 66, respectively.

Since the filters are separable, a 2D filter response can be performed by cascading two 1D filters horizontally and vertically. All filters operate at full input data rate, but the output can be subsampled in each direction by a factor k. In many cases, although not always, the filter size n and the subsampling coefficient k satisfy the following relationship.
n = 2 ^* k-1
This represents a 50% coverage overlap for the subsampled area. As an example, F3 62, which is the overall 2D response of the smallest 3 × 3 filter, is as follows:

Larger filters are described as well. Since these filters are separable, it is best to run them in two 1D steps orthogonal to each other. Each filter output is normalized by the sum of coefficients to adapt it back to the 8-bit range. Some filters, such as F3 filter 62, have a sum of weights that is a power of two. These filters can simply be implemented as a right shift rounding process of 2, so no division is required in the normalization stage. For example, the F3 filter 62 has a total 1D weight 1 + 2 + 1 = 4. This rounding division by weight can be achieved by adding 2 and then shifting to the right by 2.

Normalized result = (sum + 2) >> 2

  In general, when rounding is required, rounding is applied by adding half the divisor before performing the shift. Since the right shift performed on a binary-encoded two's complement is a value equivalent to the floor function (numerator / 2 ^ shift), by adding half the divisor, the signed numerator and sign Rounding to the nearest integer is performed for both the numerator and none.

  When the total weight of the filter is not added to a power of 2, by approximating it using multiplication by the ratio of the two numbers, a calculation-intensive division operation is avoided and the denominator is chosen to be a power of 2 .

Subsampling filters F3 / 2, F15 / 8, F31 / 16 and F63 / 32 all have 1D weights of powers of 2, 4, 64, 256 and 1024, respectively. Therefore, normalization is just a rounding right shift. The F5 filter 64 has 9 1D weights and can be approximated by multiplying 57 before rounding right shift by 9 positions. Note that the multiplication of x by 57 can be done by using shift / add / subtract operations without using variable multiplication as follows.

x ^* 57 = x << 6-x << 3 + x

  Referring to FIGS. 6A and 6B, the MX3Max unit 32 used in the Hi40 and Lo50 channels requires a maximum value in a 3 × 3 window centered on the current pixel 74 of interest. The input is an 8-bit signal. The request for maximum value is made for 9 pixels in a 3 × 3 window. This gray expansion module produces an 8-bit output consisting of the maximum pixel value 76 found within the search window boundaries. The MX3max algorithm is shown in FIG. 6B.

Referring now to FIGS. 7A and 7B, these C3 contrast modules 48 are designed to measure the amount of local contrast at the input. Contrast is defined as the difference between the maximum and minimum pixel values in the window centered on the current pixel of interest 74. The C3 contrast unit 48 uses a 3 × 3 window size centered on the pixel of interest 74 currently. The input to the contrast unit is an 8-bit signal. The contrast module produces an 8-bit monochrome output (single channel) 84. The operation of the C3 contrast unit 48 is shown in FIG. 7B. The calculation is as follows. At each pixel location, a 3 × 3 window context requires a minimum pixel value and a maximum pixel value separately. The output contrast value is determined as follows.

Contrast = maximum-minimum

Since the maximum and minimum pixel values are always between 0 and 255 for an unsigned 8-bit input signal, the contrast is guaranteed to be in the range [0. Conversion is not required.

  Three Min-Max detection modules 31, 32 and 33 are used to find peaks and valleys in the input signal. By counting the number of peaks and valleys per unit area, a measure of the local frequency is obtained. Each one of the Mx30, Hi50 and Lo40 channels uses similar MM3 units 31, 32 and 33. One difference between the three units is that each unit uses a different set of thresholds to adjust the frequency sensitivity of the corresponding channel, and Lo channel MM3 32 operates at 1/4 of the other two speeds. It is to be.

  All three units 31, 32 and 33 operate on a one-component gray source. Each unit uses a 3 × 3 window to indicate when the center pixel is an extrema (peak or valley) for its 8 neighboring pixels, and follows the following logic. The output from each Min-Max detection unit 31, 32 and 33 is a 1-bit signal indicating that the corresponding Src pixel is in the extreme state (can be extended to other color channels as well).

  The MM3 Min-Max detection structure is shown in FIG. For each pixel, the outer ring of eight pixels surrounding it (the current pixel of interest) is first analyzed. The eight outer pixels are further divided into two sets of four pixels, as shown in FIG. The partitioning of the outer ring into two sets reduces the risk of false alarms when detecting a straight line segment as a halftone (since the most commonly encountered halftone is probably a clustered dot). Useful for.

In each set, pixel values are compared between the set members 78 and 86 to determine the minimum and maximum values within each set separately.

For all (i, j) belonging to the set A, A _max = max (Aij)
For all (i, j) belonging to the set A, A _min = min (Aij)
For all (i, j) belonging to the set B, B _max = max (Bij)
For all (i, j) belonging to set B, B _min = min (Bij)

From these, the sum of the overall outer ring and min is calculated. The noise level is calculated using the sum of min and the configuration parameter of 2.

Noise = ConThr + X ^* NoiseFac / 256

The center pixel 74 value X is defined at the peak when [significantly] greater than any maximum pixel value in the set.

ならば（１）に戻る。

同様に、中心ピクセル７４値Ｘは、組のいずれかからの最小ピクセル値より［かなり］小さい場合の谷において定められる。 if

If so, return to (1).

Similarly, the center pixel 74 value X is defined in the valley when [substantially] smaller than the minimum pixel value from any of the sets.

ならば（１）に戻る。

上記の式は、３×３検出ウィンドウからの出力が１に設定される２つの条件を決め、他の全ての場合において出力は０に設定されるであろう。 if

If so, return to (1).

The above equation determines two conditions in which the output from the 3x3 detection window is set to 1, and in all other cases the output will be set to 0.

  The screen frequency / size module SEM uses one bilinear interpolation unit SCF36 and two double bilinear interpolation units DBI54. A single interpolation unit SCF 36 is applied to the sensitive frequency estimation Mx channel 30 as shown in FIG. 9A to generate the screen frequency signal SCF 70. The Hi40 and Lo50 channel DBI double interpolation unit 54 is used before combining them together to produce the screen size SCM 72.

Three interpolation modules interpolate (upsample) the signal back to the source resolution. The input signal is upsampled by a factor 32 in each direction to restore it to the original resolution. Each interpolation unit performs bilinear interpolation, essentially generating 32 ^* 32 = 1024 pixels for each original pixel. Step size bilinear interpolation is 1/32 ^nd of the original pixel grid. The following paragraphs describe the single and double interpolation units in detail.

  A single interpolation unit SCF 36 is applied to the subsampled output of the screen frequency estimation Mx channel 30. Its purpose is to restore the Mx channel 30 output to the full source resolution of the input to the screen estimation module SEM. The interpolation technique is based on 2D bilinear interpolation with 32x coefficients in each direction. After interpolation, an instantaneous screen frequency estimation signal SCF is sent to the screen cancellation module DSC.

  A block diagram of the single interpolation unit SCF is shown in FIGS. 9A and 9B. The input to the unit is a subsampled Mx channel output representing the screen frequency. For each input pixel, the unit produces 32 × 32 = 1024 output pixels. The thick line in FIG. 9A indicates a higher order output bandwidth. Both input and output are 8-bit monochrome signals. The output is an 8-bit screen frequency estimation signal Scf70.

The operation of the single interpolation unit SCF 36 is shown in FIG. 9B. Circled positions 88 and 90 indicate the position of the input pixel. Output pixels are located at grid intersections. For simplicity, FIG. 9B shows only 8x interpolation coefficients in each direction, but the actual unit requires supporting 32x coefficients in each direction. Step size bilinear interpolation is 1/32 ^nd of the original pixel grid. The details of execution are simple.

  The Hi and Lo channel DPI double interpolation unit is similar to the single interpolation unit SCF, except that there are two interpolation stages with additional mixing operations in between. One structure of the double interpolation unit is shown in FIG. The double interpolation unit operates on the three signals 94, 96 and 98 generated in each of the Hi and Lo magnitude estimation channels.

As can be seen in FIG. 10, each of the double interpolation units consists of two interpolation stages 100 and 102, respectively. The first stage includes interpolation 100 of A ₅ 94 and C ₅ 98 inputs by 2x in each direction. Interpolation 100 uses a simple bilinear interpolation technique. The A ₅ input 94 corresponds to the output of the F5 filter 64 unit. Note that the subscripts in FIG. 10 correspond to subsampling levels. The subscript 5 indicates that the signal has been subsampled 5 times by a factor of 1/2 (1/32 in total). C ₅ input 98 corresponds to the output of a 3 × 3 contrast unit. As shown in FIG. 1, both of these inputs are already subsampled by 32x coefficients in each direction. After interpolation, the A ₄ and C ₄ outputs of this first interpolation stage are subsampled by 1/16. This is the same as the subsampling level of the B ₄ input 96. It is possible to calculate the BMA ₄ is a differential signal obtained by subtracting the A ₄ from B _4. The magnitude fine mixing factor MFB ₃ generated by applying C ₄ to the MagFineBlenVsCon function 106 is multiplied by BmA ₃ . The BmAxC ₄ signal is the result of multiplying 108 BmA ₄ by MFB ₄ and shifting it 8 to the right. At 110, A ₄ is added to this to generate an Hl ₄ or LO ₄ signal, depending on the channel. The result is sent to a 16x bilinear interpolation unit 112 which generates a Lo or Hi output depending on the channel.

The MagFineBlenVsCon function 106 is a programmable function. In one embodiment, the above exemplary MagFineBlenVsCon function 106 can be simply calculated as y = (x−16) ^* 12, and the output is clamped between 0 and 192. The following equation incorporates this typical configuration value of MagFineBlenVsCon106.

BmA ₄ = B ₄ -A ₄
MFB ₄ = MagFineBlendVsCn3 (C ₄ ) = max (0, min (192, (C ₄ −16) ^* 12))
BmAxC ₄ = (BmA ₄ ^* MFB ₄ ) >> 8

The purpose of this configuration is to refine the size estimation at the location subject to change. If the frequency magnitude estimate in one of the channels appears to be stable and constant, that value is used as the mixed output. This occurs due to the C ₄ signal from the contrast measurement unit close to zero, thereby selecting the A input. However, if the magnitude estimate begins to change, the C ₄ signal increases and the B ₄ content begins to affect the mixed output. C ₄ is proportional to the contrast, which is proportional to the magnitude of the derivative of the change. Thus, the resulting magnitude estimate is biased towards the direction of change when a certain level of change is detected.

The screen size estimation module SCM 54 in FIG. 1 takes the Hi and Lo outputs of two double interpolation units as inputs. This is then added together and the contribution from each of the channels is as follows:

SCM = min (255, SCM _H + SCM _L )
here,
SCM _H = max (0, (Hi-MagHiFrqThr) ^* MagHiFrqFac)
SCM _L = max (0, (Lo-MagLoFrqThr) ^* MagLoFrqFac)

  FIG. 11 is a diagram illustrating the above equation and clipping the action of additional logic that limits the value of HTW to an acceptable range. An area indicated as “LA” 116 represents a line drawing area. As shown in FIG. 11, one particular color screen pattern can change from the position shown as LFHT to MFHT 124, HFHT 126 as its frequency changes from high to medium to low. The curve indicated by the position on the 2D plot is convex and the screen frequency cannot be distinguished by observing Lo or Hi alone.

FIG. 3 is a block diagram of a screen estimator module system. It is a figure which shows the one-dimensional filter response of various filter units. FIG. 4 shows the two-dimensional filter response of various units. FIG. 4 shows the two-dimensional filter response of various units. FIG. 4 shows the two-dimensional filter response of various units. FIG. 3 shows a typical 3 × 3 maximum module structure. FIG. 3 shows a typical 3 × 3 maximum module structure. FIG. 2 illustrates a typical 3 × 3 contrast module structure. FIG. 2 illustrates a typical 3 × 3 contrast module structure. Fig. 3 shows a minimum-maximum detection structure in a 3x3 window. It is a figure which shows a single interpolation unit. It is a figure which shows a single interpolation unit. It is a block diagram of the structure of one double bilinear interpolation unit. It is a figure which shows the estimation formula of the magnitude | size of a screen.

Explanation of symbols

28 Source Image 30 Mx Channels 31, 32, 33 Min-Max Texture Detector 36 SCF Interpolation 40 Hi Channels 41, 42, 52 Average Filter 44 MX3 Unit 46 Filter 50 Lo Channel 54 DBI Double Interpolation 56 Filter 61 SCM Size Estimation 70 Halftone frequency estimation 72 Halftone size estimation

Claims (3)

A method for obtaining an estimate of the screen frequency and size of an image signal,
(A) performing an estimation on one or more channels each exhibiting different sensitivities to give a high quality estimate for frequency and magnitude;
(B) combining one or more frequency estimates from independent channels to generate a frequency magnitude estimate;
A method consisting of: An estimation device for screen frequency and size,
(A) means for making an estimate in one or more channels each exhibiting different sensitivities and providing a high quality estimate for frequency and magnitude;
(B) means for combining one or more frequency estimates from independent channels to generate a frequency magnitude estimate;
A device consisting of An estimation device for screen frequency and size,
(A) means for making an estimate in one or more channels each exhibiting different sensitivities and providing a high quality estimate for frequency and magnitude;
(B) means for combining one or more frequency estimates from independent channels to generate a frequency magnitude estimate;
An apparatus characterized in that the channel exhibiting the highest sensitivity derives a frequency estimate.

Images