A Tree-Structured Filter Bank Coefficient Optimization Design Method Based on Dual-Cyclic Alternating Optimization

By using a dual-cycle alternating optimization method, the tree-structured filter bank is transformed into an equivalent structure, and the coefficients of each level of the subtree filter bank are optimized. This solves the problems of high optimization complexity and large reconstruction error of the tree-structured filter bank, and achieves efficient reduction of reconstruction error and preservation of frequency selectivity.

CN117236249BActive Publication Date: 2026-11-10HANGZHOU DIANZI UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202311268998.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-09-27
Publication Date
2026-11-10
Estimated Expiration
2043-09-27

AI Technical Summary

Technical Problem

Existing filter bank coefficient design optimization methods are not applicable to tree-structured filter banks, resulting in a complex and cumbersome optimization process and a large reconstruction error.

Method used

A method based on dual-loop alternating optimization is adopted to transform the tree-structured filter bank into an equivalent structured filter bank. The coefficients of each level of the subtree filter bank are optimized through iterative gradient optimization algorithm and alternating optimization idea, thereby reducing algorithm complexity and improving reconstruction error.

Benefits of technology

It significantly reduces the reconstruction error of tree-structured filter banks by several orders of magnitude, while maintaining the frequency selectivity essentially unchanged, and the optimization process is flexible and efficient.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117236249B_ABST
    Figure CN117236249B_ABST
Patent Text Reader

Abstract

This invention discloses a method for optimizing the coefficients of a tree-structured filter bank based on a dual-loop alternating optimization approach. For a two-stage, four-channel tree-structured filter bank, initial coefficients are first generated using an iterative gradient optimization algorithm. Then, it is equivalent to a four-channel uniform filter bank. A reconstruction error objective function is designed for the equivalent filter bank. The coefficients of the four-channel tree-structured filter bank are then alternately optimized with the goal of minimizing the maximum error, serving as the inner loop. Spectral shift optimization is then performed on the optimized filter, serving as the outer loop. The optimal solution obtained from both loops is output as the optimization result. This method can improve the reconstruction error by several orders of magnitude and allows for flexible changes to constraints according to design needs. It iteratively obtaining the filter coefficients reduces the algorithmic complexity of tree-structured filter bank optimization, while also having minimal impact on the frequency selectivity of the filter bank.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of digital signal processing technology and relates to the coefficient design of filter banks, specifically to a tree-structured filter bank coefficient optimization design method based on dual-cycle alternating optimization. Background Technology

[0002] Filter banks are an important component of multirate signal processing and are widely used in many fields such as speech processing, communication signal processing, image signal processing, adaptive filtering, and radar signal processing. The analysis filter bank first decomposes the signal to be processed into multiple sub-band signals, then processes each sub-band signal according to requirements, reducing the complexity of data processing operations. Finally, the synthesis filter bank reconstructs the processed signal to restore the original signal.

[0003] Besides the common modulation filter bank, another type of multichannel filter bank is the tree-structured filter bank. The advantage of the tree-structured filter bank is its flexible sub-band division, but its disadvantages include long delays and high computational cost. A tree-structured filter bank is composed of cascaded two-channel filter banks. When the two-channel filter banks satisfy the no-aliasing condition, the tree-structured filter bank also satisfies it. Based on this characteristic, the coefficient optimization problem of the tree-structured filter bank can be transformed into the coefficient optimization problem of its two-channel filter bank modules.

[0004] Existing filter bank coefficient design optimization methods directly optimize the coefficients of the prototype filter based on two indicators: reconstruction error and frequency selectivity. However, this type of optimization is not well-suited for tree-structured filter banks. Because there are many parameters to optimize, the optimization process for tree-structured filter banks is quite complex and tedious. Summary of the Invention

[0005] This invention proposes a tree structure filter bank coefficient optimization design method based on dual-cycle alternating optimization. This method is simple and efficient, and significantly improves the reconstruction error of the tree structure filter bank through optimization without significantly affecting the frequency selectivity.

[0006] The tree-structured filter bank coefficient optimization design method based on dual-cycle alternating optimization specifically includes the following steps:

[0007] Step 1: Determine the number of frequency sampling points L across the entire bandwidth. Determine the parameters of each stage of the input tree-structured filter bank, including the low-pass filter H in the two-channel filter bank of the root subtree when J=0. 00 order N 00 , passband cutoff frequency ω p00 and stopband cutoff frequency ω s00 The high-pass filter H in the root subtree two-channel filter bank 01 order N01 , passband cutoff frequency ω p01 and stopband cutoff frequency ω s01 When J=1, the low-pass filter H in the two-channel filter bank of the upper subtree 10_0 order N 10_0 , passband cutoff frequency ω p10_0 and stopband cutoff frequency ω s10_0 The high-pass filter H in the two-channel filter bank of the upper subtree 11_0 order N 11_0 , passband cutoff frequency ω p11_0 and stopband cutoff frequency ω s11_0 The low-pass filter H in the two-channel filter bank of the subtree 10_1 order N 10_1 , passband cutoff frequency ω p10_1 and stopband cutoff frequency ω s10_1 The high-pass filter H in the two-channel filter bank of the subtree 11_1 order N 11_1 , passband cutoff frequency ω p11_1 and stopband cutoff frequency ω s11_1 .

[0008] Then, based on the iterative gradient optimization algorithm, the initial solutions for the two-channel filter bank modules of each subtree in the four-channel tree structure filter bank are generated, i.e., the initial coefficients of the low-pass and high-pass filters in the two-channel filter banks of each subtree. Since the order of each subtree filter is even, its tap coefficients are even-symmetric. Therefore, the column vector set h consisting of the effective coefficients of the low-pass and high-pass filters in the two-channel filter banks of the root subtree at level J=0 is... 00 h 01 for:

[0009]

[0010] Similarly, the effective coefficients of the two-channel filter banks of the upper and lower subtrees at level J=1 are:

[0011]

[0012] The subscript x is used to distinguish between upper and lower subtrees. x = 0 indicates the two-channel filter bank module of the upper subtree, and x = 1 indicates the two-channel filter bank module of the lower subtree.

[0013] Step 2: Equivalent the two-stage four-channel tree-structured filter bank to a four-channel uniform filter bank, and calculate the full-frequency matrix of the equivalent filter for each channel. The specific steps are as follows:

[0014] s2.1, the low-pass filter H of the root subtree in the two-stage four-channel tree structure filter bank. 00 The low-pass filter H of the subtree10_0 The combined structure forms an equivalent low-pass filter H0; the root subtree's low-pass filter H... 00 The high-pass filter H of the subtree 11_0 The combined structure forms an equivalent bandpass filter H1; the root subtree's high-pass filter H... 10 and the low-pass filter H of the lower part 10_1 The combined structure forms an equivalent high-pass filter H2; the high-pass filter H of the root subtree. 10 The high-pass filter H of the subtree 11_1 The combined structure forms an equivalent bandpass filter H3.

[0015] s2.2 Calculate the passband and stopband ranges of the equivalent filters for each channel after conversion:

[0016] The passband of the equivalent low-pass filter H0 is stopband is

[0017] The passband of the equivalent bandpass filter H1 is stopband is

[0018] The passband of the equivalent high-pass filter H2 is stopband is

[0019] The passband of the equivalent bandpass filter H3 is stopband is

[0020] s2.3 Calculate the low-pass filter H in the two-channel filter bank of the root subtree of order J=0. 00 Full frequency point matrix U t_00 and high-pass filter H 01 Full frequency point matrix U t_01 :

[0021]

[0022] in:

[0023]

[0024] Let k represent the normalized frequency points, k = 1, 2, ..., L.

[0025] s2.4 Calculate the low-pass filter H in the two-channel filter banks of the upper and lower subtrees of level J=1 respectively. 10_x and high-pass filter H 11_x Full frequency point matrix U t_10_x and U t_11_x :

[0026]

[0027] in,

[0028]

[0029] s2.5 Calculate the amplitude-frequency response matrix of each filter in the two-stage four-channel tree-structured filter bank:

[0030]

[0031] The "." symbol represents vector-matrix multiplication.

[0032] s2.6. Based on the cascade identity of the decimators, determine the amplitude-frequency responses H0(ω), H1(ω), H2(ω), and H3(ω) of each channel of the equivalent filter in the equivalent four-channel uniform filter bank:

[0033]

[0034] The symbol “.*” represents dot product.

[0035] Step 3: Based on the minimization-maximization optimization criterion, solve the maximum error minimization problem. The specific steps are as follows:

[0036] s3.1 Determine the objective function for the reconstruction error of the four-channel uniform filter bank:

[0037] E R =|H2(ω+π)|*|H0(ω)|+|H3(ω+π)|*|H1(ω)|+|H0(ω+π)|*|H2(ω)|+|H1(ω+π)|*|H3(ω)| (9)

[0038] The asterisk (*) indicates scalar multiplication.

[0039] s3.2 The problem of minimizing the reconstruction error of a four-channel uniform filter bank in the maximum-minimum sense is described as follows:

[0040]

[0041] Among them, ||·|| ∞ This represents the calculation of the infinite norm, where α is a transitional variable in the solution process.

[0042] Step 4: Perform alternating optimization on the four-channel uniform filter bank, including two layers of loops. The outer loop performs transition band shifting, while the inner loop performs alternating optimization. The specific steps are as follows:

[0043] s4.1. Set the transition number r = 0, 1, 2, ..., R of the two-channel filter bank transition band shift in the root subtree of level J = 0. max R maxThis represents the limit of the migration order, with a migration amount Δr = 0.001π for each iteration. This is due to the low-pass filter H at order J = 0. 00 The passband cutoff frequency cannot be greater than the frequency center point and the high-pass filter H 01 The passband cutoff frequency cannot be lower than the frequency center point, so R is set... max :

[0044]

[0045] s4.2, Low-pass filter H in the root subtree of level J=0 after migration 00 The passband cutoff frequency and stopband cutoff frequency are expressed as High-pass filter H 01 The passband cutoff frequency and stopband cutoff frequency are expressed as follows:

[0046] s4.3 Update the passband and stopband ranges of each equivalent filter bank after each migration:

[0047] The passband of the equivalent low-pass filter H0 is [0 ω p0 The stopband is [ω]. s0 π];

[0048] The passband of the equivalent bandpass filter H1 is [ω p1_1 ω p1_2 The stopband is [0 ω]. s1_1 ]∪[ω s1_2 π];

[0049] The passband of the equivalent high-pass filter H2 is [ω p2 π], stopband is [0 ω] s2 ];

[0050] The passband of the equivalent bandpass filter H3 is [ω p3_1 ω p3_2 The stopband is [0 ω]. s3_1 ]∪[ω s3_2 π].

[0051] in:

[0052]

[0053] s4.4. Set the number of iterations i = 1, 2, ..., I for the inner-layer alternating optimization process. max I max This represents the maximum number of iterations; the trust region boundary is δ = 1. Based on the design requirements, the desired passband and stopband optimization objectives for the equivalent filters H0, H1, H2, and H3 in each channel of the four-channel uniform filter bank are given.

[0054] s4.5, Low-pass filter H in a two-channel filter bank of the root subtree of J=0. 00 The coefficients are optimized, and the change in coefficients during iterative optimization is set to Δh. 00 The convex optimization problem is defined as shown in equation (13) and solved using the CVX optimization function toolbox in Matlab software:

[0055]

[0056] Where α1 is the transition variable in the optimization process, and ε p_H0 ε p_H1 ε s_H0 ε s_H1 These are the low-pass filters H in the tree-structured filter bank. 00 The expected values ​​of the passband and stopband optimization of the equivalent filters H0 and H1 are affected.

[0057] s4.6, Low-pass filter H in the two-channel filter bank of the upper subtree of level J=1 10_0 and high-pass filter H 11_0 The coefficients are optimized, and the change in coefficients during iterative optimization is set to Δh. 10_0 and △h 11_0 The convex optimization problem is defined as shown in equation (14) and solved using the CVX optimization function toolbox in Matlab software:

[0058]

[0059] Among them, α2 is the transition variable in the optimization process.

[0060] s4.7, High-pass filter H in a two-channel filter bank of the root subtree of order J=0 01 The coefficients are optimized, and the change in coefficients during iterative optimization is set to Δh. 01 The convex optimization problem is defined as shown in equation (15) and solved using the CVX optimization function toolbox in Matlab software:

[0061]

[0062] Among them, α3 is the transition variable in the optimization process. These are the high-pass filter H in the tree-structured filter bank. 01 The expected values ​​of the passband and stopband optimization of the equivalent filters H2 and H3 are affected.

[0063] s4.8, Low-pass filter H in a two-channel filter bank for the subtree of level J=1 10_1 and high-pass filter H 11_1The coefficients are optimized, and the change in coefficients during iterative optimization is set to Δh. 10_1 and △h 11_1 The convex optimization problem is defined as shown in equation (16) and solved using the CVX optimization function toolbox in Matlab software:

[0064]

[0065] Among them, α4 is the transition variable in the optimization process.

[0066] s4.9 Variable coefficients of each filter after the i-th iteration:

[0067]

[0068] Substituting the updated coefficients into formulas (8) and (9), the reconstruction error E of the updated four-channel uniform filter bank is calculated. R (i) .

[0069] s4.10, Set the iteration termination condition for inner layer optimization: when or i>I max If the condition is met, exit the loop and save the optimization result; otherwise, update the iteration count i, i = i + 1, and return to s4.5.

[0070] s4.11 Determine the termination condition for outer transition band migration. When r > R max If the condition is met, exit the loop and compare all the saved results, outputting the smallest reconstruction error as the final result; otherwise, update the migration value r, r = r + 1, and return to s4.2.

[0071] The present invention has the following beneficial effects:

[0072] First, the complex optimization problem of tree-structured filter bank coefficients is transformed into the optimization problem of coefficients of equivalent structured filter banks. Then, using the alternating optimization approach, one filter is fixed while other filters are optimized, and the optimization is performed alternately. This method is repeated in a four-channel tree-structured filter bank, which can improve the reconstruction error by several orders of magnitude. Furthermore, the constraints can be flexibly changed according to design needs, and the filter coefficients can be obtained iteratively, reducing the algorithm complexity of tree-structured filter bank optimization. At the same time, this method has little impact on the frequency selectivity of the filter bank. Attached Figure Description

[0073] Figure 1 This is a schematic diagram of the four-channel tree-structured filter bank structure in the embodiment;

[0074] Figure 2 This is a schematic diagram of the equivalent four-channel uniform filter structure in the embodiment;

[0075] Figure 3 A flowchart of a tree-structured filter bank coefficient optimization design method based on dual-loop alternating optimization;

[0076] Figure 4 This is a comparison of the amplitude-frequency response of the four-channel uniform filter bank before and after optimization in Example 1;

[0077] Figure 5 This is a comparison chart of the reconstruction errors of the four-channel uniform filter bank before and after optimization in Example 1;

[0078] Figure 6 This is a comparison of the amplitude-frequency response of the four-channel uniform filter bank before and after optimization in Example 2;

[0079] Figure 7 This is a comparison chart of the reconstruction error of the four-channel uniform filter bank before and after optimization in Example 2. Detailed Implementation

[0080] The present invention will be further explained below with reference to the accompanying drawings;

[0081] like Figure 1 The four-channel tree-structured filter bank shown is composed of three cascaded two-channel biorthogonal filter bank modules, and each level of the two-channel biorthogonal filter bank module is different. Specifically, the low-pass filter in the two-channel filter bank of the root subtree at level J=0 is H. 00 The high-pass filter is H 01 The low-pass filter in the two-channel filter bank of the subtree at level J=1 is H. 10_0 The high-pass filter is H 11_0 The low-pass filter in the two-channel filter bank of the next subtree is H. 10_1 The high-pass filter is H 11_1 The low-pass filter H of the root subtree. 00 The low-pass filter H of the subtree 10_0 The combined structure forms an equivalent low-pass filter H0; the root subtree's low-pass filter H... 00 The high-pass filter H of the subtree 11_0 The combined structure forms an equivalent bandpass filter H1; the root subtree's high-pass filter H... 10 and the low-pass filter H of the lower part 10_1 The combined structure forms an equivalent high-pass filter H2; the high-pass filter H of the root subtree. 10 The high-pass filter H of the subtree 11_1 The combined equivalent bandpass filter H3 is as follows: Figure 2 As shown.

[0082] Frequency selectivity, i.e., the maximum ripple in the passband and stopband of each filter channel, and the reconstruction error ripple (MSE) are used as metrics to measure the performance of the optimized filter in this embodiment.

[0083] H0_pass=|||H0(ω1)|-1,|H0(ω2)|-1,…,|H0(ω p0 )|-1|| ∞ (18)

[0084] H1_pass=|||H1(ω p1_1 )|-1,|H1(ω p1_1+1 )|-1,…,|H1(ω p1_2 )|-1|| ∞ (19)

[0085] H2_pass=|||H2(ω p2 )|-1,|H2(ω p2+1 )|-1,…,|H2(ω L )|-1|| ∞ (20)

[0086] H3_pass=|||H3(ω p3_1 )|-1,|H3(ω p3_1+1 )|-1,…,|H3(ω p3_2 )|-1|| ∞ (twenty one)

[0087] H0_stop=|||H0(ω s0 )|-1,|H0(ω s0+1 )|-1,…,|H0(ω L )|-1|| ∞ (twenty two)

[0088] H1_stop=|||H1(ω1)|-1,…,|H1(ω s1_1 )|-1,|H1(ω s1_2 )|-1,…,|H1(ω L )|-1|| ∞ (twenty three)

[0089] H2_stop=|||H2(ω1)|-1,|H2(ω2)|-1,…,|H2(ω s2 )|-1|| ∞ (twenty four)

[0090] H3_stop=|||H3(ω1)|-1,…,|H3(ω s3_1 )|-1,|H3(ω s3_2)|-1,…,|H3(ω L )|-1|| ∞ (25)

[0091] MSE=|||H2(ω+π)|*|H0(ω)|+|H3(ω+π)|*|H1(ω)|+|H0(ω+π)|*|H2(ω)|+|H1(ω+π)*|H3(ω)|-1|| ∞ (26)

[0092] Example 1

[0093] This embodiment is... Figure 1 The two-channel filter bank modules in the four-channel tree structure filter bank shown are alternately optimized, and the specific steps are as follows: Figure 3 As shown.

[0094] Set filter parameters:

[0095] J = 0:N 00 =46,ω p00 =0.42,ω s00 =0.58;

[0096] N 01 =46,ω s01 =0.425,ω p01 =0.575;

[0097] J = 1:N 10_0 =48,ω p10_0 =0.41,ω s10_0 =0.59;

[0098] N 11_0 =46,ω s10_0 =0.41,ω p10_0 =0.59;

[0099] N 10_1 =46,ω p10_1 =0.41,ω s10_1 =0.59;

[0100] N 11_1 =48,ω s11_1 =0.41,ω p11_1 =0.59;

[0101] Optimize the desired outcome:

[0102]

[0103]

[0104] The metrics before and after optimization are shown in Table 1:

[0105]

[0106] Table 1

[0107] The amplitude-frequency response and reconstruction error of the four-channel uniform filter bank before and after optimization are compared as follows: Figure 4 , Figure 5 As shown.

[0108] Example 2

[0109] The filter parameters set in this embodiment are as follows:

[0110] J = 0:N 00 =60,ω p00 =0.42,ω s00 =0.58;

[0111] N 01 =60,ω s01 =0.425,ω p01 =0.575;

[0112] J = 1:N 10_0 =62,ω p10_0 =0.41,ω s10_0 =0.59;

[0113] N 11_0 =60,ω s10_0 =0.41,ω p10_0 =0.59;

[0114] N 10_1 =60,ω p10_1 =0.41,ω s10_1 =0.59;

[0115] N 11_1 =62,ω s11_1 =0.41,ω p11_1 =0.59;

[0116] Optimize the desired outcome:

[0117]

[0118]

[0119] The metrics before and after optimization are shown in Table 2:

[0120]

[0121] Table 2

[0122] The amplitude-frequency response and reconstruction error of the four-channel uniform filter bank before and after optimization are compared as follows: Figure 6 , Figure 7 As shown.

[0123] As can be seen from the amplitude-frequency response comparison chart and the reconstruction error comparison chart, after optimizing the coefficients using this method, the frequency selectivity of the filter bank is basically unaffected, and the reconstruction error of the filter bank is significantly improved. The data in Tables 1 and 2 also show that after optimizing the filter bank coefficients using this method, while maintaining the frequency selectivity essentially unchanged, the reconstruction error of the filter bank is significantly improved by 5 to 6 orders of magnitude; and the improvement in reconstruction error increases with the filter order.

Claims

1. A tree-structured filter bank coefficient optimization design method based on dual-cycle alternating optimization, characterized in that: Specifically, the following steps are included: Step 1: Generate the initial solution for the two-channel filter bank module of each subtree in the four-channel tree structure filter bank based on the iterative gradient optimization algorithm; for Low-pass and high-pass filters of the two-channel filter bank of the root subtree , Extract their effective coefficients to form a column vector set h. 00 h 01 : in, , They are respectively , The order of; against The effective coefficients of the two-channel filter banks in the upper and lower subtrees of the level are extracted to form a column vector set: Among them, subscript Used to distinguish between upper and lower subtrees This represents a two-channel filter bank in the upper subtree. This represents a two-channel filter bank in the subtree; , They are respectively The order of the low-pass filter and high-pass filter in the subtree; Step 2: Equivalent the two-stage four-channel tree structure filter bank to a four-channel uniform filter bank. Specifically, this involves converting the low-pass filter of the root subtree... and the low-pass filter of the subtree Combined into an equivalent low-pass filter Low-pass filter for root subtree High-pass filter for subtree Combined into an equivalent bandpass filter High-pass filter for root subtree and the low-pass filter of the subtree Combined into an equivalent high-pass filter High-pass filter for root subtree and the high-pass filter of the subtree Combined into an equivalent bandpass filter Calculate the full-frequency matrix of the equivalent filter for each channel after the conversion. With amplitude frequency response ; Step 3: Determine the objective function for the reconstruction error of the four-channel uniform filter bank: in" "Represents number multiplication; Based on the minimization-maximization optimization criterion, the problem of minimizing reconstruction error is described as follows: in, The calculation of the infinity norm. These are transitional variables in the solution process; Step 4: Perform alternating optimization of the four-channel uniform filter bank using a two-layer loop. The specific steps are as follows: s4.1, Settings Number of transitions in the transition band shift of the two-channel filter bank of the root subtree With each migration amount , , Indicates the maximum number of migrations: in, For low-pass filter The passband cutoff frequency; For high-pass filters The passband cutoff frequency; s4.2 Calculate the low-pass filter for each migration. With high-pass filter The passband cutoff frequency and the stopband cutoff frequency; s4.3 Update the passband and stopband ranges of each equivalent filter bank in the four-channel uniform filter bank: Equivalent low-pass filter The passband is The stopband is ; Equivalent bandpass filter The passband is The stopband is ; Equivalent high-pass filter The passband is The stopband is ; Equivalent bandpass filter The passband is The stopband is ; in: in, For low-pass filter The stopband cutoff frequency; For high-pass filters The stopband cutoff frequency; , Low-pass filters The passband cutoff frequency and stopband cutoff frequency, , High-pass filters The passband cutoff frequency and the stopband cutoff frequency; , Low-pass filters The passband cutoff frequency and stopband cutoff frequency, , High-pass filters The passband cutoff frequency and the stopband cutoff frequency; s4.4, Set the number of iterations i and the trust region boundary of the inner alternating optimization process as follows: , , Indicates the maximum number of iterations; given the equivalent filter of each channel in a four-channel uniform filter bank. The desired optimization targets for the passband and stopband; s4.5, to Low-pass filter in a two-channel filter bank of a root subtree The coefficients are optimized, and the change in coefficients during iterative optimization is set to... Solve the convex optimization problem as shown in equation (13): in, It is a transitional variable in the optimization process. These are the low-pass filters in the tree-structured filter bank. The effect of the equivalent filter The expected values ​​of the passband and stopband optimization targets; s4.6, to Low-pass filter in the two-channel filter bank of the upper subtree of the level and high-pass filter The coefficients are optimized, and the change in coefficients during iterative optimization is set to... and Solve the convex optimization problem as shown in equation (14): in, It is a transitional variable in the optimization process; s4.7, to High-pass filter in a two-channel filter bank of a root subtree The coefficients are optimized, and the change in coefficients during iterative optimization is set to... Solve the convex optimization problem as shown in equation (15): in, It is a transitional variable in the optimization process. These are the high-pass filters in the tree-structured filter bank. The effect of the equivalent filter The expected values ​​of the passband and stopband optimization targets; s4.8, to Low-pass filter in the two-channel filter bank of the subtree of the level and high-pass filter The coefficients are optimized, and the change in coefficients during iterative optimization is set to... and Solve the convex optimization problem as shown in equation (16): in, It is a transitional variable in the optimization process; s4.9, Calculation after the first... The variable coefficients of each filter after the next iteration: Calculate the reconstruction error using the updated coefficients ; s4.10, when or If the condition is met, exit the loop and save the optimization result; otherwise, update the iteration count. , Returns to s4.5; s4.11, when If the condition is met, exit the loop and compare all saved results; output the smallest reconstruction error as the final result. Otherwise, update the migration value. , , return to s4.

2.

2. The tree structure filter bank coefficient optimization design method based on dual-cycle alternating optimization as described in claim 1, characterized in that: In step 2, the passband and stopband ranges of the equivalent filters for each channel are as follows: Equivalent low-pass filter The passband is The stopband is ; Equivalent bandpass filter The passband is The stopband is ; Equivalent high-pass filter The passband is The stopband is ; Equivalent bandpass filter The passband is The stopband is ; Amplitude-frequency response of the equivalent filter for each channel They are respectively: in" " indicates dot product; The amplitude-frequency response matrices of each filter in the two-stage four-channel tree-structured filter bank are as follows: " "" indicates vector-matrix multiplication.

3. The tree structure filter bank coefficient optimization design method based on dual-cycle alternating optimization as described in claim 1, characterized in that: Use the CVX optimization function toolbox in Matlab to solve the convex optimization problem in step four.

4. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of any one of claims 1 to 3.

Citation Information

Patent Citations

  • Two-channel orthogonal mirror image filter bank design method based on iterative gradient search

    CN110061717A

  • Alternating iteration-based two-channel approximate orthogonal graph filter bank coefficient optimization method

    CN113096052A