CN1295653C - Circuit for realizing direct two dimension discrete small wave change - Google Patents

Abstract

The present invention relates to a circuit for directly implementing two-dimensional discrete wavelet transformation, which belongs to the technical field of image data processing. The present invention particularly relates to the hardware implementation of direct two-dimensional discrete wavelet transformation, and has the purpose to reduce hardware cost and system time delay. The present invention is suitable for the structural design of a very large scale integration (VLSI) of the high speed / low power direct two-dimensional discrete wavelet transformation of all filters with limited length. The present invention comprises a selector, a data serial parallel conversion interface circuit and a main conversion circuit, wherein the selector, the data serial parallel conversion interface circuit and the main conversion circuit are sequentially connected through electric signals; the main conversion circuit which is an output structure with 4 inputs and 4 outputs comprises a first horizontal filter, a filter second horizontal filter, a first vertical filter, a second vertical filter and an output coefficient regulation unit; the first horizontal filter is parallel to the second horizontal filter; the first vertical filter and the second vertical filter are arranged in parallel and are respectively connected with the first horizontal filter and the second horizontal filter; each of the parallel and the vertical filters is an output structure with 2 inputs and 2 outputs; the first vertical filter and the second vertical filter output the connection output coefficient regulation unit; one output electric signal of the output coefficient regulation unit is connected to an external memory, and the external memory outputs the electric signal to one input terminal.

Description

A kind of circuit of realizing direct two-dimensional discrete wavelet conversion

Technical field

The invention belongs to image data processing technology field, particularly the hardware of direct two-dimensional discrete wavelet conversion is realized.

Background technology

Wavelet transformation is because at the multiresolution and the zoom feature of time domain-frequency domain, thereby obtained application more and more widely in fields such as signal analysis, Flame Image Process, compression of images.And become a important component part such as new many Standard of image compression such as Joint Photographic Experts Group JPEG2000.Wavelet transformation is a kind of packing density computing, and in order to satisfy the needs that signal is handled in real time, the hardware of wavelet transformation is realized significant, and therefore, many in recent years researchers have proposed a large amount of algorithm and hardware configuration design about wavelet transformation.Implementation of wavelet can be divided into two big classes: a kind of convolution algorithm that is based on, a kind of boosting algorithm that is based on.Realize realizing having many advantages based on the wavelet transformation that promotes, reduce memory size, effectively reduce the calculated amount etc. of multiply-add operation such as realizing that original position is calculated based on traditional convolution algorithm.Utilize two-dimensional wavelet transformation can realize multiple dimensioned decomposition for image, have two kinds of methods to calculate the wavelet transformation of 2-D: a kind of is to realize by the separable method of row-column transform, utilize one-dimensional wavelet transform earlier image to be decomposed along line direction, and then along being listed as to decomposing.The shortcoming of this method is both to have needed a large-scale data transposition storer to deposit the intermediate data of one-dimensional operation, has big system delay again, and long computation period.Another kind is the inseparable two-dimensional wavelet transformation structures of direct two peacekeepings, this structure be for image be expert at and column direction carry out simultaneously, thereby the intermediate data storage device that requires of can cancellation separable implementation method, the required memory size of a large amount of minimizings, calculate simultaneously owing to 4 sub-band transformses again, so system has higher data throughput capabilities and little system delay.But, because therefore the proprietary wave filter that this implementation structure can need four correspondences will need to consume a large amount of hardware resources to realize the filtering operation of low-low pass, low-high pass, height-low pass and height-high pass respectively.Wherein, the paper of P.Wu---" Anefficient architecture for two-dimensional discrete wavelet transform, " IEEETrans.on Circuits and Systems for Video Tech., vol.11, no.4, pp.536-545,2001; The paper of F.Marino---" Two fast architectures for the direct 2-D discretewavelet transform-Signal Processing; " IEEE Trans.on Signal Processing, vol.49, no.6, pp.1248-1259,2001, and " Efficient high-speed low-power pipelinedarchitecture for the direct 2-D discrete wavelet transform; " IEEE Transactionson Circuits and Systems II:Analog and Digital Signal Processing, vol.47, no.12, pp.1476-1491,2000; Paper with T.Park---" High speed lattice based VLSIarchitecture of 2D discrete wavelet transform for real-time video signalprocessing; " Consumer Electronics, IEEE Trans.on Consumer Electronics, vol.48, no.4, the project organization of propositions such as 2002. has superior performance.But, the characteristics that its project organization one side mainly is aimed at orthogonal wavelet transformation design, not too be suitable for bi-orthogonal wavelet transformation,, thereby have higher hardware complexity and system's output time-delay on the other hand again because it adopts is traditional Wavelet Transformation Algorithm based on convolution.

Summary of the invention

According to defective that exists in the above-mentioned background technology and deficiency, the invention provides a kind of circuit of realizing direct two-dimensional discrete wavelet conversion, purpose is to reduce hardware cost and system delay, and be suitable for VLSI (very large scale integrated circuit) (VLSI) structural design of the direct two-dimensional discrete wavelet conversion of high speed/low-power of all finite length filtering devices and the VLSI structural design that adopts the direct two-dimensional discrete wavelet conversion of 9/7 small echo.

A kind of circuit of realizing direct two-dimensional discrete wavelet conversion of the present invention comprises that the serial/parallel conversion interface circuit of selector switch, data, the main transformer of electric signal connection change circuit successively; It is 4 inputs/4 export structures that main transformer changes circuit, parallel first vertical filter and second vertical filter and the regular unit of output coefficient that comprises the first parallel horizontal filter and second horizontal filter, is connected respectively with them, described level and vertical filter are 2 inputs/2 export structures, and output is connected the regular unit of output coefficient to first vertical filter with second vertical filter; Low-low frequency sub-band output the electric signal of the regular unit of output coefficient is connected to the input end that external memory storage, external memory storage export selector switch to.

Described a kind of circuit of realizing direct two-dimensional discrete wavelet conversion is characterized in that the regular unit of described output coefficient is made up of two multipliers: one of output of first vertical filter connects first multiplier, multiply by the regular factor K of yardstick therein ²Back output, two directly outputs of the output of first vertical filter; One of output of second vertical filter connects second multiplier, multiply by the regular factor 1/K of yardstick therein ²Back output, two directly outputs of the output of second vertical filter.

Described a kind of circuit of realizing direct two-dimensional discrete wavelet conversion, it is characterized in that (1) described first horizontal filter and second horizontal filter be made up of 4 selector switchs, 4 multiply-add operation devices, 5 unit delay registers and 7 pipeline registers: the first unit delay register, first pipeline register, the second multiply-add operation device, second pipeline register, the 4th unit delay register, the 3rd pipeline register, the 4th multiply-add operation device, the 7th pipeline register are connected successively; The second unit delay register, the first multiply-add operation device, the 4th pipeline register, the 3rd unit delay register, the 5th pipeline register, the 3rd multiply-add operation device, the 6th pipeline register, the 5th unit delay register are connected successively; The output that the input of the first unit delay register inserts first selector, the first unit delay register inserts the first selector and the first multiply-add operation device simultaneously, and the output of first selector also inserts the first multiply-add operation device; The input that the output of the 3rd unit delay register inserts second selector, the 3rd unit delay register inserts the second selector and the second multiply-add operation device simultaneously, and the output of second selector also inserts the second multiply-add operation device; The output that the input of the 4th unit delay register inserts third selector, the 4th unit delay register inserts third selector and the 3rd multiply-add operation device simultaneously, and the output of third selector also inserts the 3rd multiply-add operation device; The output that the input of the 5th unit delay register inserts the 4th selector switch, the 5th unit delay register inserts the 4th selector switch and the 4th multiply-add operation device simultaneously, and the output of the 4th selector switch also inserts the 4th multiply-add operation device; (2) difference of the circuit structure of described first vertical filter and second vertical filter and first horizontal filter and second horizontal filter is constituent parts chronotron in the horizontal filter is replaced with the line chronotron, the eight pipeline register output of the input of the 5th line delay register through increasing; The delay unit DU of consisting of of each line chronotron: a n cascade _n-DU ₁Select 1 selector switch mux with a n, each delay unit is respectively by N/2 ⁿUnit delay register, N/2 ⁿUnit delay register, N/2 ^N-1Unit delay register, N/2 ^N-2Unit delay register, N/2 to the last ²The series connection of unit delay register, the output of this n delay unit are connected to the selector switch input end, selector switch is exported the output as the line chronotron, and N is picture traverse and is 2 integral number power that n gets smaller or equal to log2 ^NArbitrary integer.

Described a kind of circuit of realizing direct two-dimensional discrete wavelet conversion, it is characterized in that the serial/parallel conversion interface circuit of data is made up of 4 shift register concatenation, the corresponding respectively input end that is connected to first horizontal filter and second horizontal filter of the output of each shift register, finish the serial/parallel conversion of data, thereby obtain the input of 4 data at each internal work clock.

Circuit structure of the present invention carries out the design in structured data path based on the lifting wavelet transform algorithm, adopts concurrent technique and pipelining, and embedded data boundary treatment technology, realizes parallel high speed/low-power dissipation system structure Design.The main body of circuit of the present invention comprises two horizontal filter HF1 and HF2, and two vertical filter VF1 and VF2.Horizontal filter and the work of vertical filter parallel pipeline, the utilization factor of each filter module is 100%.This structure has realized that 4 sub-band transformses carry out simultaneously, and produces corresponding 4 coefficients a clock period.Proposed to adopt the folded form structure to realize the system architecture of multilevel wavelet decomposition transform, at about N ²(1-4 ^-J)/3 clock period can be finished the J level wavelet decomposition conversion of N * N image block.

Circuit structure of the present invention is a kind of VLSI design proposal with high speed/low power capabilities.Relatively be operated in other single-input single-output system structure under the same frequency, will 4 times improve the processing power of systems.If it is constant with the processing speed that keeps system to reduce the frequency of operation of system, will effectively reduce the operating voltage of system, significantly reduce the power consumption of system, will be respectively original 1/2 and 1/16.Relatively based on the homogeneous circuit structure of convolution realization, circuit structure of the present invention will have hardware cost and system's output time-delay still less.The circuit structure that proposes has high speed/low-power consumption, rule, low characteristics such as control complexity, is highly susceptible to VLSI design realization.

Description of drawings

Fig. 1 is a circuit structure block diagram of the present invention.

Fig. 2 is the example structure figure that main transformer of the present invention changes circuit.

Fig. 3 (a) is that mode is read in the data input among Fig. 1 of the embodiment of the invention, and Fig. 3 (b) is the serial/parallel conversion interface circuit synoptic diagram of the input data among Fig. 1.F among the figure _sRepresent sample frequency, f _wRepresent the internal circuit frequency of operation.

Fig. 4 (a) is the circuit structure embodiment of the horizontal filter HF1/HF2 of Fig. 2.

Fig. 4 (b) is the circuit structure embodiment of the vertical filter VF1/VF2 of Fig. 2.

Fig. 5 is the configuration diagram of Fig. 4 (b) center line chronotron (LD).

Embodiment

The present invention is described in detail below in conjunction with accompanying drawing and embodiment.

According to the technical scheme of Fig. 1, provided the VLSI design generic structure diagram that is suitable for all finite length filtering device wavelet transformations.Use in the present embodiment be recommend in the JPEG2000 standard diminish wavelet filter-CDF (9/7) bi-orthogonal wavelet transformation, the structure of proposition not only is suitable for this type of wave filter.The circuit of Fig. 1 comprises: external memory unit EM, the input serial/parallel conversion interface circuits of data " S/P conversion ", the first horizontal filter HF1, the second horizontal filter HF2, the first vertical filter VF1 and the second vertical filter VF2.

In the present embodiment, at first according to the universal architecture of Fig. 1 proposition and the direct two-dimentional 9/7 wavelet transformation structural drawing of Fig. 2 proposition, requiring primary module is 4 inputs/4 export structures.The circuit structure of Fig. 1 and Fig. 2 comprises the horizontal filter of two concurrent workings and the vertical filter of concurrent working, and described level and vertical filter are 2 inputs/2 export structures.

Carry out data read by adopting 4 times of clocks that the view data of external memory unit is carried out " class is form in a zigzag " scanning, promptly carry out data sampling and obtain one group of input data x according to the order of even number line even column, even number line odd column, odd-numbered line even column, odd-numbered line odd column to internal operating frequencies _Ee(m, n), x _Eo(m, n), x _Oe(m, n) and x _Oo(m n), and finishes serial/parallel (S/P) conversion of data by 4 shift registers, thereby obtains the input of 4 data at each internal work clock, and its data read mode and input Data Format Transform interface circuit design are shown in Fig. 3 (a) and Fig. 3 (b).

In the circuit of Fig. 2, horizontal filter and vertical filter are finished the one-dimensional transform that follows with column direction respectively.The row that the first horizontal filter HF1 and the second horizontal filter HF2 concurrent working are carried out even number line and odd-numbered line data respectively simultaneously produces the low frequency coefficient L of two corresponding horizontal directions to conversion an internal work clock period _e(m, n) and L _o(m, n), and two high frequency coefficient H _e(m, n) and H _o(m, n), two low frequency coefficient L that horizontal filter produces _e(m, n) and L _o(m n) outputs to the input end of the first vertical filter VF1, two high frequency coefficient H of generation _e(m, n) and H _o(m n) will output to the VF2 input end of second vertical filter.Be listed as to conversion by VF1 and VF2, produce 4 subbands of image simultaneously respectively: low-low frequency (LL), low-high frequency (HL), height-low frequency (LH) and height-high frequency (HH).

In order to reduce system delay, vertical filter and horizontal filter cascade, parallel pipeline work, and be listed as to conversion to scanning along row equally.The structure of horizontal filter and vertical filter is respectively shown in Fig. 4 (a) and Fig. 4 (b).First horizontal filter and second horizontal filter are made up of 4 selector switchs, 4 multiply-add operation devices, 5 unit delay registers and 7 pipeline registers among Fig. 4 (a): the first delay register D1 of unit, the first pipeline register R1, the second multiply-add operation device PEb, the second pipeline register R2, the 4th delay register D4 of unit, the 3rd pipeline register R3, the 4th multiply-add operation device PEd, the 7th pipeline register R7 connect successively; The second delay register D2 of unit, the first multiply-add operation device PEa, the 4th pipeline register R4, the 3rd delay register D3 of unit, the 5th pipeline register R5, the 3rd multiply-add operation device PEc, the 6th pipeline register R6, the 5th delay register D5 of unit connect successively; The output that the input of the first delay register D1 of unit inserts first selector mux1, the first delay register D1 of unit inserts the first selector mux1 and the first multiply-add operation device PEa simultaneously, and the output of first selector mux1 also inserts the first multiply-add operation device PEa; The input that the output of the 3rd delay register D3 of unit inserts second selector mux2, the 3rd delay register D3 of unit inserts the second selector mux2 and the second multiply-add operation device PEb simultaneously, and the output of second selector mux2 also inserts the second multiply-add operation device PEb; The output that the input of the 4th delay register D4 of unit inserts third selector mux3, the 4th delay register D4 of unit inserts third selector mux3 and the 3rd multiply-add operation device PEc simultaneously, and the output of third selector mux3 also inserts the 3rd multiply-add operation device PEc; The output that the input of the 5th delay register D5 of unit inserts the 4th selector switch mux4, the 5th delay register D5 of unit inserts the 4th selector switch mux4 and the 4th multiply-add operation device PEd simultaneously, and the output of the 4th selector switch mux4 also inserts the 4th multiply-add operation device PEd.The structure of the project organization of vertical filter and horizontal filter is basic identical among Fig. 4 (b), and the eight pipeline register R8 output of input through increasing that just simply the delay register D1-D5 of unit in the horizontal filter is replaced with line (or row) chronotron LD1-LD5, the 5th line delay register LD5 gets final product.Because the down-sampling characteristic of wavelet transformation, the length of line delay wherein only is half of former input data line width, and the concrete structure of design line delay circuit as shown in Figure 5.In Fig. 5, each line chronotron LD1-LD5 consists of: the delay unit DU of n cascade _n-DU ₁Select 1 selector switch mux with a n, each delay unit is respectively by N/2 ⁿUnit delay register, N/2 ⁿUnit delay register, N/2 ^N-1Unit delay register, N/2 ^N-2Unit delay register, N/2 to the last ²The series connection of unit delay register, N is picture traverse and is 2 integral number power that n gets smaller or equal to log2 ^NArbitrary integer, the output of this n delay unit is connected to the output of selector switch input end, selector switch output as the line chronotron.Be output as each multiply-add operation device its first input signal and the second input signal sum multiply by promote coefficient then with the 3rd input signal addition; Be that PEa is output as: I ₃+ a (I ₁+ I ₂); PEb is output as: I ₃+ b (I ₁+ I ₂); PEc is output as: I ₃+ c (I ₁+ I ₂); PEd is output as: I ₃+ d (I ₁+ I ₂).As an embodiment, when N gets 128, n got 3 o'clock, and then the line chronotron comprises delay unit DU ₃, DU ₂, DU ₁With one 3 select 1 selector switch, wherein DU ₃Form DU by 32 unit delay registers ₂Form DU by 32 unit delay registers ₁Form by 64 unit delay registers.

Lifting decomposition algorithm by wavelet transformation, mapping obtain respective horizontal wave filter and vertical filter realize row to the structure of row to one-dimensional wavelet transform, shown in Fig. 4 (a) and Fig. 4 (b), each sub-piece is implemented in the single work clock cycle and finishes the data processing of 2 inputs/2 outputs, prediction with upgrade to promote and each step promotes arithmetic pipelining executed in parallel to improve the processing power of system.

For the multilevel wavelet decomposition transform of J 〉=2, adopt folding structure mode shown in Figure 1 to realize.Low-the low frequency coefficient of the 1st grade of conversion generation is buffered to external memory storage (EM) unit (because the advantage that the original position of lifting wavelet transform is calculated, for image compression system, this storer can be replaced by the storer of storage raw image data).After the 1st grade of conversion finished, select the data in the 1 selector switch read buffer memory to carry out the 2nd grade conversion again by one 2, this process is until till the decomposed class that requires finishes, and exports corresponding 4 subband signal LL (LL) respectively ^j, LH (LL) ^j, HL (LL) ^jAnd HH (LL) ^j(j=0～J-1, the decomposed class that the J representative is the highest).

In order to reduce the on-chip memory scale and to the visit of storer, to adopt embedded data boundary symmetry expansion technique to carry out the boundary treatment of data, it realizes selecting 1 data selector to realize by 4 * 42, sees shown in Fig. 4 (a) and Fig. 4 (b).The embedded data boundary expansion algorithm of realizing forward transform is as follows:

(1) the capable embedded data boundary symmetry extension process algorithm of one dimension to conversion:

$H^{\left(i\right)}(m,n)=\left\{\begin{array}{c}{H}^{(i-1)}(m,n)+s^i(L^{(i-1)}(m,n)+L^{(i-1)}(m,n+1)),\mathrm{for\; n}=\mathrm{0,1},...,N/2-2\\ H^{(i-1)}(m,n)+{2s}^i\×L^{(i-1)}(m,n),\mathrm{for\; n}=N/2-1\end{array}\right.,$

for?i＝1，2. (1a)

$L^{\left(i\right)}(m,n)=\left\{\begin{array}{c}{L}^{(i-1)}(m,n)+t^i(H^{\left(i\right)}(m,n)+H^{\left(i\right)}(m,n-1)),\mathrm{for\; n}=1,...,N/2-1\\ L^{(i-1)}(m,n)+{2t}^i\×H^{\left(i\right)}(m,n),\mathrm{for\; n}=0\end{array}\right.,$

for?i＝1，2. (1b)

L in the formula ⁽⁰⁾(m, n)=(m 2n) is the even column sample, H to x ⁽⁰⁾(m, n)=(m 2n+1) is the odd column sample to x.

(2) one dimension is listed as the embedded data boundary symmetry extension process algorithm to conversion:

${\mathrm{HL}}^{\left(i\right)}(m,n)=\left\{\begin{array}{c}{\mathrm{HL}}^{(i-1)}(m,n)+s^i(L{L}^{(i-1)}(m,n)+L{L}^{(i-1)}(m+1,n)),\mathrm{for\; m}=\mathrm{0,1},...,N/2-2\\ {\mathrm{HL}}^{(i-1)}(m,n)+{2s}^i\×L{L}^{(i-1)}(m,n),\mathrm{for\; m}=N/2-1\end{array}\right.,$

for?i＝1，2. (2a)

${\mathrm{LL}}^{\left(i\right)}(m,n)=\left\{\begin{array}{c}{\mathrm{LL}}^{(i-1)}(m,n)+t^i(H{L}^{\left(i\right)}(m,n)+H{L}^{\left(i\right)}(m-1,n)),\mathrm{for\; m}=\mathrm{0,1},...,N/2-1\\ {\mathrm{LL}}^{(i-1)}(m,n)+{2t}^i\×H{L}^{\left(i\right)}(m,n),\mathrm{for\; m}=0\end{array}\right.,$

for?i＝1，2. (2b)

${\mathrm{HH}}^{\left(i\right)}(m,n)=\left\{\begin{array}{c}{\mathrm{HH}}^{(i-1)}(m,n)+s^i({\mathrm{LH}}^{(i-1)}(m,n)+{\mathrm{LH}}^{(i-1)}(m+1,n)),\mathrm{for\; m}=\mathrm{0,1},...,N/2-2\\ H{H}^{(i-1)}(m,n)+{2s}^i\×{\mathrm{LH}}^{(i-1)}(m,n),\mathrm{for\; m}=N/2-1\end{array}\right.$

for?i＝1，2. (2c)

$L{H}^{\left(i\right)}(m,n)=\left\{\begin{array}{c}L{H}^{(i-1)}(m,n)+t^i({\mathrm{HH}}^{\left(i\right)}(m,n)+{\mathrm{HH}}^{\left(i\right)}(m-1,n)),\mathrm{for\; m}=\mathrm{0,1},...,N/2-1\\ L{H}^{(i-1)}(m,n)+{2t}^i\×{\mathrm{HH}}^{\left(i\right)}(m,n),\mathrm{for\; m}=0\end{array}\right.,$

for?i＝1，2. (2d)

In formula (1) and (2), s ¹=a, s ²=c; t ¹=b, t ²=d is the lifting coefficient of 9/7 wavelet transformation, LL ⁽⁰⁾(m, n)=L ⁽²⁾(2m, n), HL ⁽⁰⁾(m, n)=L ⁽²⁾(2m+1, n), LH ⁽⁰⁾(m, n)=H ⁽²⁾(2m, n), HH ⁽⁰⁾(m, n)=H ⁽²⁾(2m+1, n).

The total system internal unity is synchronous working under a single-phase clock control, and steering logic and control circuit are very simple.The present invention also analyzes the structure that proposes, and structure and other relatively more effective structure that the present invention proposes have been carried out performance relatively, comprising the parallelism wave filter structure of [1] C.Chakrabartiand M.Vishwanath (1995); [2] the direct two-dimensional structure of P.Wu and L.Chen (2001) based on polynomial expression decomposition and folding; [3] the structure of F.Marino (2000) based on inseparable two-dimensional wavelet transformation; [4] K.Andra, 4 processor structures of C.CHakrabarti and T.Acharya (2002) based on boosting algorithm; And [5] L.Liu, X.Wang, H.Meng, the recurrence pyramid structure that and et.al (2003) proposes based on " spatial array election algorithm ".Structure wherein [1], [2], [3] are based on the convolution algorithm wavelet transformation and realize, structure [4], and [5] are based on that the lifting wavelet transform algorithm proposes.Structure [1] and [2] mainly are that its filter length of choosing is K at the optimizing structure of orthogonal wavelet.Comparative result sees Table 1 in detail.

Comparative result shows that the present invention has more performance.Although used multiplier and the totalizer relatively structure of [4] and [5] are some more, other performance comprises: the memory size that requires in system delay, the sheet, processing power and to the visit of storer with control aspect such as complexity good performance is all arranged.

Table 1: performance relatively

K, L, M: filter length; J: decomposed class; The data of gray scale form are at 9/7 small echo.

Claims (5)

1. a circuit of realizing direct two-dimensional discrete wavelet conversion comprises that the serial/parallel conversion interface circuit of selector switch, data, the main transformer of electric signal connection change circuit successively; It is 4 inputs/4 export structures that main transformer changes circuit, parallel first vertical filter and second vertical filter and the regular unit of output coefficient that comprises the first parallel horizontal filter and second horizontal filter, is connected respectively with them, described level and vertical filter are 2 inputs/2 export structures, and output is connected the regular unit of output coefficient to first vertical filter with second vertical filter; Low-low frequency sub-band output the electric signal of the regular unit of output coefficient is connected to the input end that external memory storage, external memory storage export selector switch to.

2. a kind of circuit of realizing direct two-dimensional discrete wavelet conversion as claimed in claim 1 is characterized in that the regular unit of described output coefficient is made up of two multipliers: one of output of first vertical filter connects first multiplier, multiply by the regular factor K of yardstick therein ²Back output, two directly outputs of the output of first vertical filter; One of output of second vertical filter connects second multiplier, multiply by the regular factor 1/K of yardstick therein ²Back output, two directly outputs of the output of second vertical filter.

3. a kind of circuit of realizing direct two-dimensional discrete wavelet conversion as claimed in claim 1 or 2 is characterized in that: (1) described first horizontal filter and second horizontal filter are made up of 4 selector switchs, 4 multiply-add operation devices, 5 unit delay registers and 7 pipeline registers: the first unit delay register, first pipeline register, the second multiply-add operation device, second pipeline register, the 4th unit delay register, the 3rd pipeline register, the 4th multiply-add operation device, the 7th pipeline register are connected successively; The second unit delay register, the first multiply-add operation device, the 4th pipeline register, the 3rd unit delay register, the 5th pipeline register, the 3rd multiply-add operation device, the 6th pipeline register, the 5th unit delay register are connected successively; The output that the input of the first unit delay register inserts first selector, the first unit delay register inserts the first selector and the first multiply-add operation device simultaneously, and the output of first selector also inserts the first multiply-add operation device; The input that the output of the 3rd unit delay register inserts second selector, the 3rd unit delay register inserts the second selector and the second multiply-add operation device simultaneously, and the output of second selector also inserts the second multiply-add operation device; The output that the input of the 4th unit delay register inserts third selector, the 4th unit delay register inserts third selector and the 3rd multiply-add operation device simultaneously, and the output of third selector also inserts the 3rd multiply-add operation device; The output that the input of the 5th unit delay register inserts the 4th selector switch, the 5th unit delay register inserts the 4th selector switch and the 4th multiply-add operation device simultaneously, and the output of the 4th selector switch also inserts the 4th multiply-add operation device; (2) difference of the circuit structure of described first vertical filter and second vertical filter and first horizontal filter and second horizontal filter is constituent parts chronotron in the horizontal filter is replaced with the line chronotron, the eight pipeline register output of the input of the 5th line delay register through increasing; The delay unit DU of consisting of of each line chronotron: a n cascade _n-DU ₁Select 1 selector switch mux with a n, each delay unit is respectively by N/2 ⁿUnit delay register, N/2 ⁿUnit delay register, N/2 ^N-1Unit delay register, N/2 ^N-1Unit delay register, N/2 to the last ²The series connection of unit delay register, the output of this n delay unit are connected to the selector switch input end, selector switch is exported the output as the line chronotron, and N is picture traverse and is 2 integral number power that n gets smaller or equal to log2 ^NArbitrary integer.

4. a kind of circuit of realizing direct two-dimensional discrete wavelet conversion as claimed in claim 1 or 2, it is characterized in that the serial/parallel conversion interface circuit of data is made up of 4 shift register concatenation, the corresponding respectively input end that is connected to first horizontal filter and second horizontal filter of the output of each shift register, finish the serial/parallel conversion of data, thereby obtain the input of 4 data at each internal work clock.

5. a kind of circuit of realizing direct two-dimensional discrete wavelet conversion as claimed in claim 3, it is characterized in that the serial/parallel conversion interface circuit of data is made up of 4 shift register concatenation, the corresponding respectively input end that is connected to first horizontal filter and second horizontal filter of the output of each shift register, finish the serial/parallel conversion of data, thereby obtain the input of 4 data at each internal work clock.

Images