JP4217408B2 - Filter processing device - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a filter processing apparatus, and more particularly, to a filter processing apparatus that performs wavelet transform on image data and inversely transforms wavelet transform coefficients into image data.
[0002]
[Prior art]
An image, particularly a multi-valued image, contains a great deal of information, and there is a problem that the amount of data is enormous when storing and transmitting the image. For this reason, when storing and transmitting images, high-efficiency coding that reduces the amount of data by removing the redundancy of the images or changing the contents of the images to such an extent that deterioration in image quality is difficult to visually perceive. Used.
[0003]
For example, in JPEG recommended by ISO and ITU-T as an international standard encoding method for still images, image data is subjected to discrete cosine transform (DCT) for each block (8 pixels × 8 pixels) and converted to DCT coefficients. Later, each coefficient is quantized and further entropy encoded to compress the image data. However, in this method, since DCT and quantization are performed for each block, so-called block distortion may appear at the boundary of each block of the decoded image.
[0004]
On the other hand, JPEG2000 is being studied as a new international standard encoding method for still images. In JPEG2000, wavelet transform is proposed as a conversion process performed before quantization. The wavelet transform has a feature that it is difficult to visually understand the degradation of the decoded image because the input data is continuously processed instead of being processed in units of blocks as in the current JPEG.
[0005]
In the wavelet transform used in JPEG2000, the conversion process can be efficiently performed with a small amount of calculation by performing a process using a method called a lifting mechanism.
[0006]
FIG. 12 shows a signal flow in the forward lifting mechanism, and FIG. 13 shows a signal flow in the reverse lifting mechanism. Α, β, γ, and δ in the figure are called lifting coefficients.
[0007]
First, the operation of the lifting mechanism shown in FIG. 12 will be described.
[0008]
Input pixels in the order of input X₀, X₁, X₂, X_Three, X_Four, X_Five, ... in order. The input pixels are classified into an even pixel series and an odd pixel series by the classification unit 201, and a pixel X whose index is even from one output terminal (upper side in FIG. 12) of the classification unit 201.₀, X₂, X_Four... (ie X_2n) From the other output terminal (lower side in FIG. 12)₁, X_Three, X_Five... (ie X_{2n + 1}) Is output.
[0009]
In the first-stage lifting process, the even pixel series is multiplied by a lifting coefficient α, and the multiplication result of two consecutive even pixels is added to the pixels in the odd pixel series located at the center of the two pixels.
[0010]
This can be expressed as a generalized expression as follows.
[0011]
D_{2n + 1} = X_{2n + 1} + Α · X_2n+ Α · X_{2n + 2}          ... (1)
In the second stage lifting process, the newly obtained odd pixel series D₁, D_Three, D_Five,... Are multiplied by a lifting coefficient β, and the multiplication result of two consecutive odd pixels is added to the pixels in the even pixel series located at the center of the two pixels.
[0012]
This can be expressed as a generalized expression as follows.
[0013]
E_{2n + 2}= X_{2n + 2}+ Β · D_{2n + 1}+ Β · D_{2n + 3}        ... (2)
In the third stage lifting process, the lifting coefficient γ is used in the same way as the first stage, and in the fourth stage lifting process, the lifting coefficient δ is used as in the second stage. Expressions representing the contents of the third and fourth lifting processes are as follows.
[0014]
H_{2n + 1} = D_{2n + 1} + Γ · E_2n   + Γ · E_{2n + 2}          ... (3)
L_{2n + 2}= E_{2n + 2} + Δ · H_{2n + 1} + Δ · H_{2n + 3}          (4)
In FIG. 12, K normalizes the wavelet coefficient, but is not particularly relevant in explaining the essence of the present invention, so the explanation is omitted here.
[0015]
If the normalization process is ignored, the H level obtained by the 3rd and 4th stage lifting processes can be obtained._n, L_n Respectively correspond to a high-frequency transform coefficient and a low-frequency transform coefficient.
[0016]
Next, the signal flow of the reverse lifting mechanism shown in FIG. 13 will be briefly described. First, in response to the normalization process in the forward lifting mechanism, the inverse coefficient is multiplied, and then the four-stage lifting process is performed. The processing contents of each stage are summarized below and expressed by an expression.
[0017]
(First stage) E2_{n + 2}= L_{2n + 2} -Δ · H_{2n + 1} -Δ · H_{2n + 3}      ... (5)
(Second stage) D_{2n + 1} = H_{2n + 1} − Γ ・ E_2n   − Γ ・ E_{2n + 2}      ... (6)
(3rd stage) X2_{n + 2} = E_{2n + 2} -Β · D_{2n + 1} -Β · D_{2n + 3}      ... (7)
(4th stage) X2n + 1 = D2n + 1-α · X2n-α · X2n + 2 (8)
The above formulas (5), (6), (7), and (8) are obtained by shifting the formulas (4), (3), (2), and (1), respectively.
[0018]
The lifting mechanism shown in FIGS. 12 and 13 is expressed from another viewpoint to the lifting lattice structure shown in FIGS. In the figure, □ represents input data, ◯ represents a grid point (or grid point data calculator), and an arrow from ◯ represents a flow of grid point data. In these drawings, the basic processing in the lifting mechanism (the processing of the above formulas (1) to (8)) and new data obtained by the processing are associated with one grid point.
[0019]
In the forward-direction lifting grid structure shown in FIG. 14, one grid point data is calculated using any one of the equations (1) to (4).
[0020]
In the lifting grid structure in the reverse direction shown in FIG. 15, one grid point data is calculated by any one of the equations (5) to (8).
[0021]
In an ordinary filter, one output is calculated each time one piece of data is input. However, as understood from the lifting grid structure of FIG. 14, in the lifting calculation processing, only two new data are prepared. Two data outputs are possible.
[0022]
For example, X₈In the input data up to, the output data is L_Four, H_Five Can only calculate up to. Next X₉However, there is no grid point data that can be newly calculated. But X_Ten D is newly provided by D₉, E₈, H₇, L₆Can be computed. In addition, X₁₁, X₁₂H is the output data only after the two input data are prepared.₉, L₈Can be computed.
[0023]
As described above, in the filtering process based on the lifting operation, every time two new input data are prepared, two outputs (conversion coefficients) can be calculated. Similarly, in the inverse transformation process shown in FIG. 15, it is understood that two restoration data can be calculated every time two transformation coefficients are prepared.
[0024]
Further, when applied to the inverse conversion process in the vertical direction, two lines of conversion coefficients of 9 types of low and high frequencies are input in the horizontal scan order, so that the restored data for 2 lines is converted into the horizontal scan order. At the same time and output.
[0025]
This is a major difference between the filtering processing (wavelet transform) based on the lifting operation and the filtering processing that is not so.
[0026]
As described above, the low-frequency and high-frequency transform coefficients output in pairs from the wavelet transform unit that performs horizontal (or vertical) wavelet transform processing are performed in the vertical (or horizontal) wavelet transform processing, respectively. The next wavelet transform unit performs processing, and horizontal and vertical two-dimensional wavelet transform processing is performed by the two conversion processes.
[0027]
The processing unit for performing the two-dimensional wavelet transform has conventionally been configured as shown in FIG. 16 as disclosed in JP-A-10-283342. In the figure, reference numeral 501 denotes a horizontal one-dimensional DWT (Discrete Wavelet Transform) processing unit (hereinafter referred to as “horizontal DWT processing unit”), and 503 and 505 denote vertical one-dimensional DWT processing units (hereinafter referred to as vertical DWT). (Referred to as “processing unit”) 511 and 513 are buffers.
[0028]
The horizontal DWT processing unit 501 receives and processes raster scan data scanned in the horizontal direction, and outputs two conversion coefficients of a horizontal low range and a high frequency for each process. The buffers 511 and 513 divide the low-frequency and high-frequency conversion coefficients output from the horizontal DWT processing unit 501 and store the conversion coefficients for one horizontal line, respectively.
[0029]
On the other hand, in the vertical DWT processing units 503 and 505, immediately after the vertical one-dimensional wavelet transform process is performed, transform coefficients for a plurality of lines used for the transform process are stored in the internal buffer. When the conversion process is completed, the conversion coefficients for two lines are unnecessary. When conversion coefficients for two new lines are input from the horizontal DWT processing unit 501 and the buffers 511 and 513, the next vertical wavelet conversion process is performed. It becomes possible. The vertical DWT processing units 503 and 505 perform arithmetic processing using the input new set of conversion coefficients, and output two low-frequency conversion coefficients and a high-frequency conversion coefficient, respectively.
[0030]
In this way, by performing two types of wavelet transform processing, horizontal and vertical, the vertical DWT processing unit 503 causes LL (vertical-low frequency, horizontal-low frequency), HL (vertical-high frequency, horizontal-low frequency). 2) conversion coefficients from the vertical DWT processing unit 505, two conversion coefficients of LH (vertical-low range, horizontal-high range) and HH (vertical-high range, horizontal-high range) are provided. Is output.
[0031]
In the configuration of FIG. 16, the horizontal DWT processing unit 501 can be operated at an operation rate of 100% if two data are input every cycle. On the other hand, the two vertical DWT processing units 503 and 505 are paused without performing any processing while the conversion coefficients for the next horizontal line are stored in the buffers 511 and 513, respectively. When the conversion coefficient of the second line is input from the horizontal DWT processing unit 501 to the vertical DWT processing units 503 and 505, the conversion coefficients stored in the buffers 511 and 513 are read to use the conversion coefficient for two lines. Perform wavelet transform processing.
[0032]
Accordingly, during the period in which the two vertical DWT processing units 503 and 505 operate, the horizontal DWT processing unit 501 sets the conversion coefficient for the second line among the two line conversion coefficients newly input to the vertical DWT processing units 503 and 505. It is the same as the period of processing and outputting. That is, each of the two vertical DWT processing units operates at an operation rate of 50%.
[0033]
[Problems to be solved by the invention]
As described above, in the conventional two-dimensional wavelet transform process, two transform processing units are required to be arranged in the subsequent stage among the transform processes in the horizontal direction and the vertical direction. That is, only one conversion processing unit is required in the previous stage to perform the same amount of processing. However, since two conversion processing units are used in the subsequent stage, hardware resources cannot be effectively used, and the circuit scale increases. There was a problem.
[0034]
The present invention has been made in view of the above problems, and an object thereof is to more effectively utilize hardware resources and realize a two-dimensional wavelet transform processing device with a smaller hardware configuration.
[0035]
[Means for Solving the Problems]
  In order to achieve the above object, a filter processing apparatus of the present invention comprises the following arrangement. That is,
  A filter device for wavelet transforming image data using a plurality of arithmetic units having a lifting grid structure,
  The first or second horizontal or vertical direction of the image data to be wavelet transformed is defined as the Y axis, the second direction orthogonal to the first direction is defined as the X axis, and the pixel position x of the X coordinate. , Pixel data at the pixel position y of the Y coordinate is defined as Px, y, and a row of pixel data at the same Y coordinate position in the image data is defined as a row
  In the i-th processing cycle (i = 0, 1, 2,...), A set of two pixel data adjacent in the first direction {P_{x + i, y}, P_{x + i, y + 1}} And the pixel data P input one line before_{x + i, y-1}And at least the pixel data P_{x + i, y-1}Using the calculation result of the calculation unit calculated based on_{x + i, y}, P_{x + i, y + 1}} From one low-frequency conversion coefficient L_iAnd one high frequency conversion coefficient H_iA first direction wavelet transform means for outputting
  Frequency transform coefficient L generated in two processing cycles of the wavelet transform means in the first direction_i, H_iAnd L_{i + 1}, H_{i + 1}Each time is input, the four frequency conversion coefficients are rearranged, and the low frequencyStrangeConversion coefficient pair {L_i, L_{i + 1}} And a set of high frequency conversion coefficients {H_i, H_{i + 1}}, Each of which is output in units of one processing cycle;
  A set of low frequency conversion coefficients {L_i, L_{i + 1}} In one processing cycle, the low frequency conversion coefficient L input one line before_i-1And at least the low frequency conversion coefficient L_i-1And the input set {L using the calculation result of the calculation unit calculated based on_i, L_{i + 1}}, One low and low frequency conversion coefficient LL and one low and high frequency conversion coefficient LH are output, and a set of high frequency conversion coefficients {H_i, H_{i + 1}} In one processing cycle, the high frequency conversion coefficient H input one line before_i-1And at least the high-frequency conversion coefficient H_i-1Using the calculation result of the calculation unit calculated based on_i, H_{i + 1}}, Second directional wavelet transform means for outputting one high / low frequency transform coefficient HL and one high / high frequency transform coefficient HH.
[0041]
  Also, a preferred one of the present inventionAspectAccording to the aboveWavelet transform in the second directionThe means is an FIR filter.
[0051]
DETAILED DESCRIPTION OF THE INVENTION
Preferred embodiments of the present invention will be described below in detail with reference to the accompanying drawings.
[0052]
FIG. 1 shows a grid point data calculation unit that performs calculation at each grid point shown in FIG. 14, and FIG. 2 shows a multistage connection of each grid point data calculation unit shown in FIG. 1 for performing a lifting calculation of filter processing. Shows the configuration.
[0053]
In FIG. 1, 601 and 603 are terminals for inputting two data, 607 is a terminal for outputting calculated lattice point data, 621 is a buffer for storing input data from the terminal 603, and 609 is an output data of the buffer 621 externally. 611 is an adder that adds the output data of the buffer 621 and the input data from the terminal 603, and 613 is one of coefficients C (α, β, γ, δ) added to the addition result of the adder 611. 615 is an adder for adding the multiplication result of the multiplier 613 to the input data located at the center of the three data used for the calculation.
[0054]
First, the outline of the calculation method in the embodiment of the present invention will be briefly described with reference to FIGS. In the following description, data output from each grid point in FIG. 14 is also referred to by the same reference name as the grid point.
[0055]
For example, 9 input data X₀, X₁, X₂, X_Three, X_Four, X_Five, X₆, X₇, X₈10 grid point data (D₁, D_Three, D_Five, D₇, E₂, E_Four, E₆, H_Three, H_Five, L_Four) To calculate the low frequency conversion coefficient L_FourAnd high-frequency conversion coefficient H_FiveCan be output.
[0056]
Next X₉, X_TenWhen two pieces of data are newly added as input data, the low-frequency transform coefficient L is calculated by calculating 10 grid point data in the same manner as described above.₆And high-frequency conversion coefficient H₇Can be output, but X₉, X_TenIf the grid point data calculated before is input, it is necessary to newly calculate D₉, E₈, L₇, L₆No more than four.
[0057]
In order to use the previously calculated grid point data, a medium for storing and holding the grid point data after the calculation is required, which is the buffer 621 in FIG.
[0058]
Only the buffer in the uppermost grid point data calculation unit 701 in FIG. 2 is used to hold previously input data, not the previously calculated grid point data, but in other grid point data calculation units. This buffer is used to hold previously calculated grid point data. This buffer has a minimum size of 1 and no upper limit.
[0059]
X₀To X₈The processing using the data of is already finished, and the low-frequency conversion coefficient L₆And high-frequency conversion coefficient H₇In the uppermost grid point data calculation unit 701, new input data X₉, X_Ten Only two of these are input. In the grid data unit 701, D₉Input data X₉And X_TenIn addition to the data X required for this calculation₈Is output from the buffer 621 in FIG. This X₈X in the previous process₈Is stored in the buffer 621 when it is input from the terminal 603.
[0060]
The grid point data calculation unit 701 calculates the calculated D₉And output X from buffer 621₈Are output from the terminals 607 and 609 to the outside of the unit, and sent to the next grid point data calculation unit 702.
[0061]
The grid point data calculation unit 702 receives the input D₉And X₈Using E₈, But another data D necessary for this calculation₇Is output from the buffer 621 in the unit 702. This D₇Is stored in the buffer 621 when it is input from the terminal 603 in the previous processing. And the calculated E₈And output D from buffer 621₇Are output from the terminals 607 and 609 to the outside of the unit, and sent to the next grid point data calculation unit 703.
[0062]
The lattice point data calculation units 703 and 704 also perform the same processing as described above. As a result, the high frequency conversion coefficient H is output from the arithmetic unit 703.₇However, from the arithmetic unit 704, the low frequency conversion coefficient L₆Are output respectively.
[0063]
Thereafter, each time two pieces of new data are input to the arithmetic unit 701, high-frequency and low-frequency conversion coefficients are output from the arithmetic units 703 and 704.
[0064]
When the buffer 621 shown in FIG. 1 has a single-stage register 1701 for storing data for one pixel as shown in FIG. 17, it is possible to perform the same horizontal wavelet transform processing as before. As shown in FIG. 18, in the case of two-stage registers 1801 and 1802 each storing data for one pixel, two kinds of signals are processed by alternately processing two kinds of signals as will be described later. It is possible to perform the wavelet transform process. The data flow in FIG. 18 is first input to the register 1801 from the top of the figure, shifted to the register 1802 at the next timing, and further output from the bottom of the figure at the next timing. Further, as shown in FIG. 19, when the line memory 1901 is configured to store data for one line of the image, the vertical wavelet transform process can be performed.
[0065]
(First embodiment)
Next, a two-dimensional wavelet transform process using the wavelet transform processing unit having the above-described configuration in the first embodiment of the present invention will be described.
[0066]
In the first embodiment of the present invention, the two-dimensional wavelet transform process is realized by dividing the one-dimensional transform process with different directions into two stages, and at that time, the two-stage transform process is performed. In between, 2 × 2 data rotation processing is performed.
[0067]
FIG. 3 shows the configuration of the two-dimensional wavelet transform processing apparatus in the first embodiment. In the figure, reference numeral 901 denotes a vertical one-dimensional DWT processing unit (hereinafter referred to as “vertical DWT processing unit”), 903 denotes a data rotation unit that performs rotation processing of 2 × 2 data, and 905 denotes a horizontal one-dimensional unit. DWT processing unit (hereinafter referred to as “horizontal DWT processing unit”). As shown in FIG. 2, each of the vertical DWT processing unit 901 and the horizontal DWT processing unit 905 has a configuration in which the arithmetic units shown in FIG. 1 are cascade-connected, and the vertical DWT processing unit 901 includes four arithmetic units. The line memory 1901 (FIG. 19) as the buffer 621 (FIG. 1) and the horizontal DWT processing unit 905 have the two-stage registers 1801 and 1802 (FIG. 18) as the buffer 621 in each of the four-stage arithmetic units. And a horizontal one-dimensional wavelet transform process.
[0068]
Pixel data of one pixel from the end of each line of pixel data for two lines, ie, two pixels arranged in the vertical direction, is sequentially supplied to the vertical DWT processing unit 901 from a memory or line buffer (not shown).
[0069]
The vertical DWT processing unit 901 uses the newly received vertical two-pixel data and the internal buffer 621, that is, the pixel data input at the previous timing stored in the line memory 1901 (FIG. 19). The vertical low-frequency conversion coefficient L_vAnd high-frequency conversion coefficient H_vAre output one by one. By using a line memory that stores image data for one line, it is possible to perform calculations between three pixels that are lined up vertically, with the pixel data input one line before and the newly received vertical two-pixel data. It becomes.
[0070]
The processing of the vertical DWT processing unit 901 will be described in detail.
[0071]
FIG. 20 is a diagram illustrating a state in which any two adjacent arithmetic units 2000a and 2000b among the four arithmetic units constituting the vertical DWT processing unit 901 are connected in cascade. Here, description will be made as an example in which the arithmetic units 701 and 702 in FIG. 2 are connected. That is, 2000a and 2000b correspond to the arithmetic units 701 and 702, respectively. In the arithmetic units 2000a and 2000b, the multiplication coefficients C of the multipliers 2013a and 2013b are α and β, respectively. Further, the two outputs 2007a and 2009a of the arithmetic unit 2000a are connected to the two inputs 2001b and 2003b of the arithmetic unit 2000b. Furthermore, as described above, the buffer 621 is configured by the line memories 2021a and 2021b, respectively. The remaining two-stage arithmetic units 703 and 704 (not shown) have the same configuration.
[0072]
The data flow in FIG. 20 will be described below. Here, the subscript number of the pixel data represents the vertical position of the pixel data.
[0073]
From one input terminal 2001a of the arithmetic unit 2000a, certain pixel data X out of one line of even-numbered pixel data._2n-1Is input to the adder 2015a. From the other input terminal 2003a, out of one line of pixel data of odd lines, X_2n-1Data X in a vertical relationship with_2nIs input to the adder 2011a and simultaneously to the line memory 2021a.
[0074]
In the line memory 2021a, pixel data X that is exactly perpendicular to the input pixel_2n-2Are output to the output terminal 2009a and the adder 2011a. Pixel data X_2n-2Is data input to the line memory 2021a before the line period. In the adder 2011a, X_2nAnd X_2n-2Are output to the multiplier 2013a. Where X_2n-2Indicates that the position on the image is X_2nX in the data two lines before_2nAnd the pixel positions are in a vertical relationship, and they are added. In the multiplier 2013a, the result of multiplying the coefficient C = α by α (X_{2 n}+ X_2n-2) Is output to the adder 2015a. In the adder 2015a, the multiplication result from the multiplier 2013a and the input X from the input terminal 2001a._2n-1And add D_2n-1= X_2n-1+ Α (X_2n+ X_2n-2) And output to the output terminal 2007a.
[0075]
By calculating the above calculation for all the lines of the pixel data, the calculation for calculating the odd pixel series D corresponding to the above-described equation (1) is performed.
[0076]
In the arithmetic unit 2000b, it is possible to obtain the even data series E corresponding to the above equation (2) using the odd data series obtained above as an input. The calculation in 2000b is exactly the same as that in 2000a except that the input series 2003b is D and the coefficient C = β, and detailed description thereof is omitted.
[0077]
In this way, it is possible to obtain the low-frequency transform coefficient and the high-frequency transform coefficient in the vertical direction using the arithmetic units connected in cascade in four stages.
[0078]
The two output conversion coefficients L_v, H_vIs input to the data rotation unit 903. When the rotation unit 903 receives two low-frequency and high-frequency conversion coefficients, the rotation unit 903 rearranges them and outputs them in the order of two low-frequency conversion coefficients and two high-frequency conversion coefficients. This input / output relationship is conceptually shown in FIG.
[0079]
FIG. 4A shows two sets of data divided by input units, and FIG. 4B shows two sets of data divided by output units. Comparing (a) and (b), it can be seen that (b) is obtained by rotating (a) 90 degrees to the right.
[0080]
Since the rearranged two low-frequency transform coefficients and two high-frequency transform coefficients are transform coefficients that are continuous in the horizontal direction, if this is processed by the horizontal DWT processing unit 905, horizontal wavelet transform can be performed. it can.
[0081]
In the first embodiment, a data rotation unit is provided to rearrange data, and the rearranged low-frequency and high-frequency two types of data are alternately displayed in the horizontal DWT processing unit 905 as low-frequency and high-frequency. Multiplex processing.
[0082]
Multiplexing processing is performed by alternately processing two types of signals when the buffer 621 (FIG. 1) of each grid point data calculation unit of the horizontal DWT processing unit 905 includes two register stages 1801 and 1802 (FIG. 18). The horizontal wavelet transform is performed alternately on the two types of signals, low and high. This process will be described in more detail.
[0083]
In the current processing cycle, data for low-frequency conversion coefficient calculation is stored in the top 1802 of the two-stage registers 1801 and 1802 that are buffers 621 of each grid point data calculation unit, and high-frequency conversion coefficient calculation is performed in the subsequent stage 1801. It is assumed that data for use is stored. At this time, the data for high-frequency conversion coefficient calculation is only connected to the register 1802 ahead, and no calculation unit refers to the data. This is equivalent to the absence of high-frequency conversion coefficient calculation data, and all the arithmetic units are in a state of processing low-frequency conversion coefficients.
[0084]
Therefore, in this cycle, two low-frequency conversion coefficients L of the conversion coefficients shown in FIG._v1And L_v2To process the data (here, data for low-frequency conversion coefficient calculation) output from the first register 1802 of the two-stage register together with the low-frequency conversion coefficient, and output the processing results (LL, LH) To do. Then, in the next cycle, the data for the high-frequency conversion coefficient stored in the subsequent register 1801 moves to the top of the two-stage register 1802, and the input L_v2Alternatively, data used for processing in the current cycle, which is input from the previous grid point data calculation unit, is taken into the subsequent stage of the register 1801.
[0085]
In the next cycle, the order of the low-frequency transform coefficients and the high-frequency transform coefficients stored in the two-stage registers 1801 and 1802 is reversed, and all the arithmetic units are in a state of processing the high-frequency transform coefficients. Therefore, in the next cycle, the rotation unit 903 receives two high-frequency conversion coefficients H_v1And H_v2Is input, processed, and the processing result (HL, HH) is output.
[0086]
The processing of the horizontal DWT processing unit 905 will be described in detail.
[0087]
FIG. 21 is a diagram illustrating a state in which any two adjacent arithmetic units 2100a and 2100b among the four arithmetic units constituting the horizontal DWT processing unit 905 are connected in cascade. Here, description will be made as an example in which the arithmetic units 701 and 702 in FIG. 2 are connected. That is, 2100a and 2100b correspond to the arithmetic units 701 and 702, respectively. In the arithmetic units 2100a and 2100b, the multiplication coefficients C of the multipliers 2113a and 2113b are α and β, respectively. Also, the two outputs 2107a and 2109a of the arithmetic unit 2100a are connected to the two inputs 2101b and 2103b of the arithmetic unit 2100b. Further, as described above, the buffer 621 includes the two-stage registers 2121a and 2123a, 2121b and 2123b, respectively. The remaining two-stage arithmetic units 703 and 704 (not shown) have the same configuration.
[0088]
The data flow in FIG. 21 will be described below. Here, the subscript number of the conversion coefficient represents the horizontal position of the conversion coefficient.
[0089]
First, as is apparent from the above, a series L in which low-frequency transform coefficients and high-frequency transform coefficients output from the rotation unit 903 are alternately arranged at the input terminal 2101a of the arithmetic unit 2100a._2m-3, H_2m-3, L_2m-1, H_2m-1, L_{2m + 1}, H_{2m + 1}, ... are entered. Similarly, the input terminal 2103a has L_2m-2, H_2m-2, L_2m, H_2m, L_{2m + 2}, H_{2m + 2}, ... are entered. The two input timings are L_2m-1And L_2mAre input at the same time.
[0090]
Low frequency conversion coefficient L from input terminal 2101a_2m-1Is input to the adder 2115a and at the same time the low-frequency conversion coefficient L from the input terminal 2103a._2mIs input to the adders 2111a and 2121a.
[0091]
L_2m-1,_{L 2m}Is input to the register 2121a at the previous timing from the input terminal 2103a._2m-2Is stored. At this time, the low-frequency conversion coefficient L input from the input terminal 2103a to the register 2123a at the previous timing is also displayed._2m-2Is stored.
[0092]
Therefore, in the adder 2111a, L inputted from the input terminal 2103a_2mIs the low-frequency conversion coefficient L stored in the register 2123a._2m-2And added to the multiplier 2113a.
[0093]
The multiplier 2113a multiplies it by a coefficient C = α to obtain a result α (L_2m+ L_2m-2) Is output to the adder 2115a. The adder 2115a receives the multiplication result from the multiplier 2113a and the input L from the input terminal 2101a._2m-1And DL consisting of low-frequency conversion series_2m-1= L_2m-1+ Α (L_2m+ L_2m-2) And output to the output terminal 2107a. The output value L of the register 2123a is supplied to the other output terminal 2109a._2m-2Is output.
[0094]
At the next timing, each of the input terminals 2101a and 2101b has a high-frequency conversion coefficient H._2m-1, H_2mIs entered. At the same time, the stored data is shifted in the two-stage registers 2121a and 2123a, respectively, and the register 2121a has L_2mThe register 2123a has H_2m-2Is stored. Therefore, the calculation result of the adder 2115a at this timing is DH composed of a high-frequency conversion coefficient._2m-1= H_2m-1+ Α (H_2m+ H_2m-2) And output to the output terminal 2107a. The other output terminal 2109a is connected to H_2m-2Is output.
[0095]
By obtaining the above calculation for all lines of pixel data, an operation for alternately obtaining the pixel conversion data series D corresponding to the above-described equation (1) for the low frequency and the high frequency is performed.
[0096]
In the arithmetic unit 2100b, the pixel series obtained above is used as an input, and the pixel series EL_2m-1= L_2m-1+ Β (D_2m-2+ D_2mOr EH_2m-1= H_2m-1+ Β (D_2m-2+ D_2m) Can be obtained alternately for low and high frequencies. The calculation in 2100b is exactly the same as that in 2100a except that the input series 2103b is D and the coefficient C = β, and detailed description thereof is omitted.
[0097]
By repeating the above processing, the low-frequency and high-frequency two types of signals are subjected to horizontal wavelet transform processing using arithmetic units connected in four stages in cascade to obtain transform coefficients LL, LH, HL, and HH, respectively. be able to.
[0098]
As described above, according to the first embodiment, by providing a rotation unit, two-dimensional wavelet transform processing can be performed by using one vertical and one horizontal one-dimensional wavelet transform processing units one by one. Therefore, the hardware configuration can be reduced.
[0099]
(Second Embodiment)
In the second embodiment, the order of the horizontal wavelet transform process and the vertical wavelet transform process in the first embodiment is exchanged. As shown in FIG. 5, the vertical DWT processing unit 901 is arranged on the output side of the horizontal DWT processing unit 905, and a new rotation unit 1101 is provided in front of the horizontal DWT processing unit 905.
[0100]
Assume that the order of pixel data input from a memory or line buffer (not shown) is the same as that in the first embodiment. That is, pixel data of two lines of pixel data from the end, that is, pixel data of two pixels arranged in the vertical direction is input to the rotation unit 1101 in parallel.
[0101]
The rotation unit 1101 rearranges the two lines of parallel data into data of two pixels with alternate lines and outputs the data. This is the same as the process of rotating the data by 90 degrees. FIG. 6 is a diagram conceptually showing the input / output relationship of the rotation unit 1101. In FIG. 6, the subscript U indicates the upper line of the two lines (upper line), L indicates the lower line (lower line), and 1 and 2 indicate the order of the pixels. As shown in FIG. 6A, by rotating the input pixel data by 90 degrees to the right, the pixel data for two pixels in the same row are alternately output.
[0102]
In this way, the pixel data for two pixels that are alternately output for each line is input to the horizontal DWT processing unit 905, and horizontal wavelet transform processing corresponding to each input data is performed. Since the two lines of data are alternately input, two types of signals are alternately wavelet transformed. Therefore, as described in the first embodiment, each grid point data in the horizontal DWT processing unit 905 is used. The buffer 621 (FIG. 6) of the arithmetic unit is composed of two stages of registers.
[0103]
The transform coefficients output by horizontal wavelet transform are a low-frequency / high-frequency transform coefficient of a certain line and a low-frequency / high-frequency transform coefficient of the next line, and are alternately output. These are further input to the rotation unit 903 and rearranged as shown in FIG.
[0104]
As a result, two low-frequency conversion coefficients and two high-frequency conversion coefficients from the two lines are alternately output from the rotation unit 903.
[0105]
Since the two conversion coefficients are arranged in the vertical direction, if these two conversion coefficients are input to the vertical DWT processing unit 901, the two conversion coefficients input and the data held in the internal buffer 621 are used. , Vertical wavelet transform processing can be performed.
[0106]
Each grid point data calculation unit of the vertical DWT processing unit 901 has a line memory as a buffer 621 and holds conversion coefficients for two lines. Note that two lines of conversion coefficients, one line for the low-frequency conversion coefficient and one line for the high-frequency conversion coefficient, are stored, but the number of conversion coefficients is half the number of original line data. Therefore, the capacity of the line memory may be one line. The line memory stores every other low-frequency transform coefficient and high-frequency transform coefficient so that the low-frequency transform coefficient and the high-frequency transform coefficient can be processed alternately.
[0107]
As described above, according to the second embodiment, the same effect as that of the first embodiment can be obtained.
[0108]
If the order of pixel data input from the memory or line buffer (not shown) to the wavelet transform processing device is changed so that data of two pixels are alternately input, the first rotation unit 1101 is not necessary.
[0109]
(Third embodiment)
In the first and second embodiments, the horizontal DWT processing unit 905 is based on the configuration shown in FIG. 2, but may be configured to be used in a general FIR filter. In this case, the low-frequency conversion coefficient and the high-frequency conversion coefficient are calculated by separate arithmetic units.
[0110]
Here, it is assumed that the wavelet transform filter has a linear phase, that is, wavelet transform filter coefficients are symmetric.
[0111]
FIG. 7 shows a configuration of a two-dimensional wavelet transform processing apparatus in the third embodiment. 7 except the data input unit 1401 shown in FIG. 9 and the filter calculation unit 1402 shown in FIG. 8 are the same as those described with reference to FIG. 5 in the second embodiment. is there.
[0112]
FIG. 8 is a schematic diagram showing the configuration of the filter calculation unit 1402. In the figure, 1201 is filter input data, 1202 is an adder for adding two input data having the same filter multiplication coefficient, 1203 is a coefficient multiplier, and 1204 and 1205 are adders for summing the multiplication results.
[0113]
Next, the configuration of the entire wavelet transform processing unit when horizontal wavelet transform processing is performed using the filter calculation unit 1402 shown in FIG. 8 will be described.
[0114]
Although it is necessary to input the same number of pixel data as the number of taps of the filter to the filter calculation unit 1402, unlike the configuration shown in FIG. 1, since there is no buffer inside the conversion processing unit, data writing to the buffer, Pixel data can be input without worrying about reading. In other words, even when a plurality of types of data conversion processes are alternately performed, there is no need to worry about the state inside the filter calculation unit 1402.
[0115]
In order to alternately convert a plurality of types of data, it is necessary to simultaneously switch all input pixel data. For example, if the number of filter taps for obtaining the transform coefficient is 9, nine data must be switched simultaneously. Since it is inefficient to receive such data directly from the memory, a new data input unit 1401 corresponding to the data supply unit 1401 is also required for the part that supplies data to the filter operation unit 1402.
[0116]
Since pixel data obtained from a CCD or a scanner is generally generated in a raster scan order, the pixel data generated in the raster scan order will be received and processed below. A data input unit 1401 shown in FIG. 9 corresponds to pixel data generated in the raster scan order. In the figure, 1301 is a terminal for inputting pixel data in the raster scan order, 1303 is a first line memory having a capacity for one line, and 1305 is a second line memory having a capacity of half of one line. , 1307 and 1309 are shift registers in which nine registers are connected, and 1311 is a selector that switches nine pixel data of two systems.
[0117]
Hereinafter, a specific operation of the data input unit 1401 will be described.
[0118]
First, input pixel data for the first line input from the terminal 1301 of the data input unit 1401 shown in FIG. 7 is stored in the first line memory 1303. Considering the processing cycle (period) of the wavelet transform processing unit which is an important component of the present invention as a reference, the overall processing balance is best when two pixel data are input during one cycle. Here, it is assumed that pixel data is input at such a rate.
[0119]
Next, while storing the input pixel data of the second line in the second line memory 1305, the data is read out from the first line memory 1303 and the second line memory 1305 one pixel at a time in each cycle, and the shift register 1307 is read. , 1309 to send data. If reading from the second line memory 1305 is delayed by one cycle rather than starting reading data from both line memories 1303 and 1305 at the same time, it is convenient for the reason described later.
[0120]
Since the second line memory 1305 stores two pixel data in one cycle and outputs one pixel data, only half of the data remains when all the pixel data of the second line has been stored. Absent. Therefore, the capacity of the second line memory 1305 only needs to be able to store half the data of one line.
[0121]
The pixel data taken into the shift registers 1307 and 1309 are alternately switched by the selector 1311 every cycle and processed by the filter operation unit 1402 shown in FIG. The phase of the pixel data in the shift register 1309 is one cycle later than that of the shift register 1307, but the timing at which the selector 1311 selects the two shift registers 1307 and 1309 is also shifted by one cycle. Thus, the selector output has the same phase data. Based on the nine data of the same phase for the two lines output in this way, the filter operation unit 1402 calculates and outputs the wavelet transform coefficients for the low frequency band and the high frequency band.
[0122]
Since the output timing of the conversion coefficient is the same as that of the second embodiment, if the conversion coefficient is rearranged by the rotation unit 903 and then converted by the vertical DWT processing unit 901, the conversion coefficient output timing is the same as that of the second embodiment. A transform coefficient that has been two-dimensional wavelet transformed at the same timing can be obtained.
[0123]
(Fourth embodiment)
Here, a configuration in which two line memories used on the input side of the filter calculation unit 1402 in the third embodiment are moved to the output side of the filter calculation unit 1402 will be described. However, although the total capacity of the line memories is the same as that of the third embodiment, each line memory having a capacity of 1/2 line is divided into three.
[0124]
The wavelet transform process in the horizontal direction and the pixel data in the raster scan order are compatible with each other, and the pixel data can be directly input to the horizontal DWT processing unit for processing. The horizontal DWT processing unit may be a filter calculation unit having the configuration shown in FIG. 8, or may be a conversion processing unit having the configuration shown in FIG.
[0125]
FIG. 10 is a block diagram illustrating a configuration of a wavelet transform processing device according to the fourth embodiment. In the configuration of FIG. 10, pixel data in the raster scan order is received, and the horizontal DWT processing unit 1500 immediately performs horizontal wavelet transform processing. As in the third embodiment, pixel data is input by two pixels per cycle.
[0126]
The low-frequency and high-frequency conversion coefficients, which are the processing results of the 2n-th line by the horizontal DWT processing unit 1500, are stored in 1501 and 1503, respectively. Further, among the processing results of the 2n + 1th line, the high frequency conversion coefficient is stored in 1505, but the low frequency conversion coefficient is not stored anywhere and is immediately sent to the vertical DWT processing unit 901 via the selector 1511. In accordance with this, the low-frequency transform coefficient one line before is read from the line memory 1501 and sent to the vertical DWT processing unit 901 together with the other coefficient of the two low-frequency transform coefficients arranged vertically.
[0127]
Thereby, the vertical DWT processing unit 901 can continuously convert the low-frequency conversion coefficient in the horizontal direction. During this time, all the horizontal high-frequency conversion coefficients of the (2n + 1) th line are stored in the line memory 1505.
[0128]
When the pixel data input for the 2n + 1 line is completed, the processing of the low-frequency conversion coefficients for the 2n and 2n + 1 lines is also completed. The data is read and sent to the vertical DWT processing unit 901 via the selector 1511 to perform conversion processing.
[0129]
Just as the processing of the low frequency conversion coefficient is performed in synchronization with the input of the pixel data of the 2n + 1 line, the processing of the high frequency conversion coefficient is performed in synchronization with the input of the pixel data of the 2n + 2 line. The pixel data of the (2n + 2) th line input during the processing of the high-frequency transform coefficient is subjected to the horizontal wavelet transform process, and the processing result is stored in the same place as the 2nth line. The line memory 1503 for storing the high-frequency conversion coefficient has the same ratio of the data to be output and the data to be stored. Therefore, the amount of data to be held does not change and the old data is updated to new data so that it is updated to new data. .
[0130]
As described above, transform coefficients subjected to two-dimensional wavelet transform can be obtained. However, the output order of transform coefficients is different from the conventional ones, and the low-frequency and high-frequency transform coefficients for image data for two lines are each pixel. Rather than alternately outputting lines each time data is processed, the horizontal DWT processing unit 1500 outputs low-frequency and high-frequency conversion coefficients for each line. In this way, by providing a line buffer that temporarily holds the low-frequency and high-frequency transform coefficients output for each line, one vertical wavelet and one horizontal one-dimensional wavelet transform processing unit can be used. Since a two-dimensional wavelet transform process can be performed, the hardware configuration can be reduced.
[0131]
(Fifth embodiment)
The first to fourth embodiments are premised on data input of 2 pixels per conversion processing cycle, but the fifth embodiment is suitable for data input at a rate of 1 pixel per conversion processing cycle. The configuration of the wavelet transform processing apparatus will be described.
[0132]
Simply reducing the data input to one pixel per cycle of the conversion process results in only an operation rate of 50% for each conversion processing unit and not effective use of hardware resources.
[0133]
Therefore, in the fifth embodiment, by performing both the vertical wavelet transform process and the horizontal wavelet transform process in one transform processing unit, the operation rate of the transform processing unit is set to 100%, and the hardware is effectively used. .
[0134]
The configuration of the fifth embodiment is shown in FIG. In the figure, 1601 is a selector for selecting and outputting one system from 2-input 2-system data, 1603 is a horizontal / vertical DWT processing unit, and 1605 is a data rotation unit that performs rotation processing of 2 × 2 data. is there.
[0135]
As in the first embodiment, from the memory or line buffer (not shown), one pixel from the end of each line of the two line data, that is, two vertical pixels, is sent to the DWT processing unit 1603 via the selector 1601. input.
[0136]
On the other hand, the DWT processing unit 1603 processes the input data in the vertical conversion mode.
[0137]
However, only four pixel data are input continuously in two cycles, and data input is stopped in the next two cycles. Since four pixel data are input in a total of four cycles, as described above, the data input rate is one pixel per cycle.
[0138]
The conversion processing unit 1603 outputs four sets of low-frequency conversion coefficients and high-frequency conversion coefficients in the vertical direction by processing the four pixel data input in the two-cycle period.
The conversion coefficients are rearranged by the rotation unit 1605 via the switches 1606 and 1607 as in the first embodiment, and two low-frequency conversion coefficients and two high-frequency conversion coefficients are sequentially selected. 1601 is input.
[0139]
In the two-cycle period in which the data input is stopped, the conversion coefficients rearranged by the rotation unit 1605 are selected by the selector 1601 and sent to the vertical / horizontal DWT processing unit 1603. In the horizontal conversion mode, the vertical / horizontal conversion processing unit 1603 performs wavelet conversion processing of two sets of two conversion coefficients input in succession for two cycles in the horizontal direction, and outputs them via switches 1606 and 1607. .
[0140]
The vertical / horizontal DWT processing unit 1603 basically has the configuration shown in FIG. 2, but the buffer in each arithmetic unit has a line memory corresponding to vertical conversion and a two-stage corresponding to horizontal conversion. The above-described operation can be performed by switching the register according to the conversion mode.
[0141]
The same applies even if two horizontal pixels of 2-line data are alternately input, the input data is processed in the horizontal conversion mode first, and the conversion coefficients rearranged in the rotation unit are then processed in the vertical conversion mode. The result can be obtained.
[0142]
In all of the first to fifth embodiments, only the forward wavelet transform process has been described. However, the reverse wavelet transform process differs only in that some multiplication operations are changed to subtraction, and the multiplication coefficient is different. Since the configuration of 2 and the upper configuration are the same in the forward direction and the reverse direction, the present invention can also be applied to the wavelet transform processing in the reverse direction.
[0143]
【The invention's effect】
As described above, according to the present invention, it is possible to realize a two-dimensional wavelet transform processing apparatus with a smaller hardware configuration by more effectively using hardware resources.
[Brief description of the drawings]
FIG. 1 is a diagram showing a configuration of an arithmetic unit in an embodiment of the present invention.
FIG. 2 is a block diagram showing a configuration for performing a lifting operation for filter processing in which the operation units of FIG. 1 are connected in multiple stages.
FIG. 3 is a block diagram illustrating a configuration of a two-dimensional wavelet transform processing device according to the first embodiment of the present invention.
4 is a diagram conceptually showing the operation of the rotating unit of FIG. 3;
FIG. 5 is a block diagram illustrating a configuration of a two-dimensional wavelet transform processing device according to a second embodiment of the present invention.
6 is a diagram conceptually showing the operation of the rotating unit of FIG. 5;
FIG. 7 is a block diagram illustrating a configuration of a two-dimensional wavelet transform processing device according to a third embodiment of the present invention.
8 is a diagram illustrating a configuration of an FIR type filter calculation unit illustrated in FIG. 7;
9 is a diagram showing a configuration of a data input unit shown in FIG.
FIG. 10 is a block diagram illustrating a configuration of a two-dimensional wavelet transform processing device according to a fourth embodiment of the present invention.
FIG. 11 is a block diagram illustrating a configuration of a two-dimensional wavelet transform processing device according to a fifth embodiment of the present invention.
FIG. 12 is a diagram illustrating a conventional configuration when a 9 × 7 filter process is realized by a lifting operation.
FIG. 13 is a diagram illustrating a conventional configuration when a 9 × 7 inverse filter process is realized by a lifting operation.
FIG. 14 is a diagram showing a lifting grid structure expressing a lifting operation of 9 × 7 filter processing;
FIG. 15 is a diagram showing a lifting grid structure expressing a lifting operation of 9 × 7 inverse filter processing;
FIG. 16 is a block diagram showing a configuration of a conventional two-dimensional wavelet transform processing device.
17 is a diagram showing that the buffer shown in FIG. 1 is composed of one stage of registers.
FIG. 18 is a diagram showing that the buffer shown in FIG. 1 is composed of two stages of registers.
FIG. 19 is a diagram showing that the buffer shown in FIG. 1 is composed of a line memory.
20 is a diagram illustrating a state in which any two adjacent arithmetic units of the four arithmetic units included in the vertical DWT processing unit illustrated in FIG. 3 are cascade-connected.
FIG. 21 is a diagram illustrating a state in which any two adjacent arithmetic units are cascade-connected among the four arithmetic units included in the horizontal DWT processing unit illustrated in FIG.
[Explanation of symbols]
503, 505, 901 Vertical DWT processing unit
501, 905 Horizontal DWT processing unit
511, 513 buffer
611, 615 adder
613 multiplier,
621 buffer
701-704 Lattice point data calculation unit
903, 1101, 1605 2 × 2 rotation unit
1311, 1511, 1601 selector
1303, 1305, 1501, 1503, 1505 Line memory
1401 Data input part
1402 Filter calculation unit
1603 Horizontal / Vertical DWT Processing Unit

Claims (2)

A filter device for wavelet transforming image data using a plurality of arithmetic units having a lifting grid structure,
The horizontal or vertical first direction of the image data to be wavelet transformed is the Y axis, the second direction orthogonal to the first direction is the X axis, and the pixel position x of the X coordinate , Pixel data at the pixel position y of the Y coordinate is defined as Px, y, and the line of pixel data at the same Y coordinate position in the image data is defined as a row,
In the i-th processing cycle (i = 0, 1, 2,...), Two sets of pixel data {P _{x + i, y} , P _{x + i, y + 1} } adjacent in the first direction are input. And input using the pixel data P _{x + i, y-1} input one row before and the calculation result of the calculation unit calculated based on at least the pixel data P _{x + i, y-1} First-direction wavelet transform means for outputting one low-frequency transform coefficient L _i and one high-frequency transform coefficient H _i from the set {P _{x + i, y} , P _{x + i, y + 1} }; ,
Each time the frequency transform coefficients L _i , H _i and L _{i + 1} , H _{i + 1} generated in two processing cycles of the wavelet transform means in the first direction are input, the four frequency transform coefficients are arranged. instead, low-frequency variable換係number set _{{L i, L i + 1} } and, a sorting means for outputting the set _{{H i, H i + 1} } of the high-frequency transform coefficients of each one treatment cycles,
In one processing cycle in which a set of low frequency conversion coefficients {L _i , L _{i + 1} } is input from the rearranging means, the low frequency conversion coefficient L _i-1 input one line before and at least the low frequency conversion coefficient One low-low frequency conversion coefficient LL and one low-high frequency conversion coefficient from the input set {L _i , L _{i + 1} } using the calculation result of the calculation unit calculated based on L _i-1 LH is output, and in one processing cycle in which the set of high-frequency conversion coefficients {H _i , H _{i + 1} } is input from the rearranging means, the high-frequency conversion coefficient H _i-1 input one line before and at least the relevant One high / low frequency conversion coefficient HL and one high / high frequency are input from the input set {H _i , H _{i + 1} } using the calculation result of the calculation unit calculated based on the high frequency conversion coefficient H _i−1. Second directional wavelet transform means for outputting a transform coefficient HH. Filtering device according to symptoms.   The filter processing apparatus according to claim 1, wherein the wavelet transform unit in the second direction is an FIR filter.

Images