📡Advanced Signal Processing Unit 5 Review

QMF banks split a signal into frequency subbands and then reconstruct it from those subbands. They're central to subband coding, multirate processing, and wavelet transforms because they provide a structured way to decompose and perfectly reconstruct signals across multiple frequency bands.

A two-channel QMF bank is the most common configuration: two analysis filters split the signal, and two synthesis filters put it back together. The core design challenge is choosing filter coefficients that cancel aliasing, avoid amplitude distortion, and preserve phase linearity. Methods like Johnston's and Jain-Crochiere's approach this as an optimization problem. QMF banks also generalize to cosine-modulated filter banks, multidimensional signals, and tree structures used in wavelet transforms.

Perfect reconstruction in QMF banks

Perfect reconstruction means the output is an exact (possibly delayed) copy of the input, with no distortion or aliasing artifacts. Three conditions must hold simultaneously for this to work.

Alias cancellation

Downsampling in the analysis stage creates spectral copies that overlap, producing aliasing. When the subband signals are upsampled and filtered in the synthesis stage, these aliased components must cancel exactly. In a two-channel system, this is enforced by the quadrature mirror relationship between the filters: the highpass analysis filter $H_1(z)$ is set to $H_0(-z)$ , so that aliasing terms introduced by one channel are equal and opposite to those introduced by the other.

No amplitude distortion

Even after alias cancellation, the combined magnitude response of the analysis-synthesis system can introduce gain variations across frequency. For perfect reconstruction, the overall transfer function must be allpass (flat magnitude). Formally, this requires:

$|H_0(e^{j\omega})|^2 + |H_1(e^{j\omega})|^2 = 1 \quad \text{for all } \omega$

This is the power complementary condition. In practice, classic QMF designs achieve this only approximately, which is why the distinction between perfect and near-perfect reconstruction matters.

Linear phase distortion

A linear phase response means every frequency component is delayed by the same amount, so the output is simply a time-shifted version of the input with no phase distortion. This requires symmetric (or antisymmetric) FIR filter coefficients in both the analysis and synthesis stages. Linear phase is especially important in audio and image applications where phase distortion causes perceptible artifacts.

Two-channel QMF banks

The two-channel case is the building block for all larger QMF structures. The input signal passes through a lowpass filter $H_0(z)$ and a highpass filter $H_1(z)$ , each followed by downsampling by 2. The synthesis side upsamples by 2 and applies synthesis filters $G_0(z)$ and $G_1(z)$ , then sums the results.

Analysis filters

$H_0(z)$ is a lowpass filter that extracts the lower half of the spectrum
$H_1(z) = H_0(-z)$ is its quadrature mirror image, a highpass filter that extracts the upper half
Together they form a complementary pair: what one passes, the other rejects

The mirror relationship $H_1(z) = H_0(-z)$ is what gives QMF banks their name. In the frequency domain, this means $H_1(e^{j\omega}) = H_0(e^{j(\omega - \pi)})$ , so the highpass response is a frequency-shifted copy of the lowpass response.

Synthesis filters

The synthesis filters must undo the analysis and cancel aliasing. For the standard two-channel QMF bank:

$G_0(z) = H_0(z)$
$G_1(z) = -H_1(z)$

The sign flip on $G_1(z)$ is what produces the alias cancellation. After upsampling and filtering, the two channels are summed to reconstruct the signal. The overall distortion transfer function becomes:

$T(z) = \frac{1}{2}\left[H_0^2(z) - H_1^2(z)\right]$

For perfect reconstruction, $T(z)$ must be a pure delay.

Polyphase representation

Polyphase decomposition rewrites each filter in terms of its even- and odd-indexed coefficients. For a filter $H(z)$ :

$H(z) = E_0(z^2) + z^{-1}E_1(z^2)$

where $E_0$ and $E_1$ are the polyphase components. The advantage is that filtering and downsampling can be interchanged using the noble identities, so you filter at the lower sampling rate. This cuts the computation roughly in half. The entire two-channel QMF bank can be expressed as a polyphase matrix, which also simplifies the perfect reconstruction analysis.

Design of QMF banks

Designing a QMF bank means finding filter coefficients that satisfy (or closely approximate) perfect reconstruction while also meeting specs like stopband attenuation and transition bandwidth.

Johnston's method

Johnston's method is a widely used approach for two-channel QMF design:

Start with a prototype lowpass FIR filter $H_0(z)$
Define an objective function that penalizes both stopband energy and deviation from the power complementary condition
Optimize the filter coefficients (typically using nonlinear optimization) to minimize this objective
Derive $H_1(z) = H_0(-z)$ from the optimized lowpass filter

Johnston published tables of optimized coefficients for various filter lengths (8, 12, 16, 24, 32 taps), which are still commonly referenced. These filters achieve near-perfect reconstruction with good stopband attenuation, though exact PR is not guaranteed.

Jain-Crochiere method

This is an iterative alternating optimization approach:

Initialize with a reasonable set of filter coefficients
Fix the analysis filters and optimize the synthesis filters to minimize reconstruction error
Fix the synthesis filters and optimize the analysis filters
Repeat until convergence

The method is flexible because you can incorporate different constraints at each step (e.g., filter length, stopband specs). It tends to converge to good solutions but can be sensitive to initialization.

Lattice structures

Lattice structures implement QMF banks using a cascade of two-port sections, each parameterized by a single rotation angle. Their advantages:

Guaranteed stability regardless of coefficient quantization
Low sensitivity to finite word-length effects compared to direct-form implementations
Modular structure where adding a section increases the filter order without redesigning existing sections
Perfect reconstruction can be structurally enforced, meaning it holds exactly even with quantized coefficients

Cosine modulated filter banks

Cosine modulated filter banks generalize the two-channel QMF idea to $M$ channels. All $M$ analysis filters are generated from a single prototype lowpass filter $p(n)$ by cosine modulation:

$h_k(n) = p(n) \cos\left[\frac{\pi}{M}\left(k + \frac{1}{2}\right)\left(n - \frac{N-1}{2}\right) + \theta_k\right]$

where $k = 0, 1, \ldots, M-1$ and $N$ is the filter length. This means you only need to design one prototype filter, and the modulation generates the entire filter bank.

Pseudo-QMF banks

Pseudo-QMF banks relax the perfect reconstruction requirement to gain design freedom. The prototype filter is optimized primarily for frequency selectivity (sharp transition bands, high stopband attenuation) rather than exact PR. Aliasing between adjacent channels is kept small but not perfectly cancelled. For many applications, the residual error is below perceptible thresholds, making pseudo-QMF banks a practical choice when filter performance matters more than mathematical exactness.

Near perfect reconstruction

Near-perfect reconstruction (NPR) filter banks achieve reconstruction error on the order of $-60$ to $-100$ dB or better. The prototype filter is optimized to jointly minimize both the reconstruction error and the stopband energy. NPR designs are sufficient for most audio and image coding applications, where quantization noise from the coder itself far exceeds the tiny reconstruction error of the filter bank.

Applications of QMF banks

Alias cancellation, Angular Parameterization of Orthonormal Wavelet Filter Banks in Power Quality: A Method for ...

Subband coding

Subband coding splits a signal into frequency bands, then allocates bits to each band based on its energy or perceptual importance. QMF banks provide the splitting and reconstruction. High-energy or perceptually important bands get more bits; quiet bands get fewer. This is the foundation of MPEG audio (MP3), which uses a 32-band pseudo-QMF bank, and JPEG 2000, which uses a wavelet-based QMF decomposition for images.

Multirate signal processing

QMF banks perform frequency-selective decimation and interpolation. Instead of a single anti-aliasing filter followed by downsampling, a QMF bank splits the signal into subbands that can each be processed at a reduced rate. Applications include sample rate conversion (e.g., 48 kHz to 44.1 kHz audio) and efficient narrowband filtering within a wideband signal.

Wavelet transforms

The discrete wavelet transform (DWT) is implemented by cascading two-channel QMF banks in a tree structure. At each level, the lowpass (approximation) output is fed into another two-channel QMF bank, producing a multiresolution decomposition. The filters used must satisfy specific regularity conditions to generate valid wavelets. Daubechies wavelets, for example, are derived from maximally flat QMF bank designs. Applications include signal denoising, compression, and feature extraction.

QMF banks vs other filter banks

Conjugate quadrature filters (CQF)

CQF banks (also called Smith-Barnwell filter banks) achieve exact perfect reconstruction using filters that satisfy:

$H_1(z) = z^{-(N-1)} H_0(-z^{-1})$

This is a different mirror relationship than the classic QMF condition $H_1(z) = H_0(-z)$ . The CQF formulation allows exact PR with FIR filters of any even length, at the cost of giving up linear phase. CQF banks are sometimes called "orthogonal" filter banks because the analysis and synthesis operations form an orthogonal transform.

Lapped orthogonal transforms (LOT)

LOTs use basis functions that overlap across block boundaries, unlike the block transforms (DCT) used in JPEG. This overlap eliminates the blocking artifacts that appear at high compression ratios. LOTs provide perfect reconstruction and can be viewed as a special case of cosine-modulated filter banks with specific constraints. They're used in image and video coding where visual quality at block boundaries is critical.

Multidimensional QMF banks

Multidimensional QMF banks extend subband decomposition to 2D (images), 3D (video), and beyond.

Separable QMF banks

Separable designs apply 1D QMF banks independently along each dimension. For a 2D image, you filter along rows first, then along columns (or vice versa). This produces four subbands: LL (lowpass-lowpass), LH, HL, and HH. Separable banks are simple to implement and analyze, and they're the standard approach in JPEG 2000's wavelet decomposition.

Non-separable QMF banks

Non-separable designs use 2D (or higher) filters that account for correlations across dimensions simultaneously. They can achieve better frequency partitioning of the 2D spectrum, particularly along diagonal directions where separable banks have poor selectivity. The trade-off is significantly more complex design and implementation. Non-separable banks are used in specialized applications like directional image analysis.

Tree-structured QMF banks

Tree structures cascade QMF banks hierarchically, typically decomposing only the lowpass subband at each level. This produces a dyadic frequency decomposition with finer resolution at low frequencies and coarser resolution at high frequencies, matching the behavior of the DWT. The tree structure is also the basis for wavelet packet decompositions, where any subband (not just the lowpass) can be further decomposed, giving adaptive frequency resolution.

Computational complexity of QMF banks

Polyphase implementation

Direct implementation of a two-channel QMF bank with length- $N$ filters requires $2N$ multiplications per input sample (filtering before downsampling). Polyphase implementation moves the filtering after downsampling, reducing this to $N$ multiplications per input sample. The steps are:

Decompose each filter into its polyphase components
Downsample the input first (split into even and odd samples)
Apply the polyphase filters at the lower rate
Combine outputs to form the subband signals

This 2:1 reduction in computation extends to $M$ -channel banks, where the savings factor is $M$ .

Fast algorithms

Beyond polyphase, further speedups exploit the structure of cosine-modulated filter banks. The cosine modulation can be computed using a DCT (or modified DCT), and the prototype filtering can use FFT-based fast convolution. The fast extended lapped transform (ELT) algorithm combines these ideas. For an $M$ -channel bank with prototype length $N = 2M$ , the complexity drops from $O(MN)$ to $O(M \log M)$ per block, making real-time implementations feasible even for large $M$ .

Limitations and challenges in QMF banks

Finite word length effects

Fixed-point implementations represent filter coefficients and signal samples with limited precision. This introduces:

Coefficient quantization error: the implemented filter differs from the designed filter
Roundoff noise: arithmetic operations accumulate rounding errors through the filter bank
Overflow: intermediate values may exceed the representable range

These effects degrade reconstruction quality. Lattice structures and noise-shaping techniques help mitigate them.

Sensitivity to coefficient quantization

Classic direct-form QMF implementations are particularly sensitive to coefficient quantization because the perfect reconstruction conditions depend on precise coefficient values. Even small perturbations can break alias cancellation. Lattice and lifting-based implementations are structurally less sensitive, which is why they're preferred in hardware and fixed-point software implementations.

Design trade-offs

Every QMF bank design involves balancing competing objectives:

Filter length vs. frequency selectivity: longer filters give sharper transitions but increase latency and computation
Perfect reconstruction vs. stopband attenuation: exact PR constrains the filter design, often limiting achievable stopband rejection
Computational complexity vs. performance: polyphase and fast algorithms reduce cost but may introduce structural constraints
Linear phase vs. minimum phase: linear phase requires symmetric filters (longer for a given selectivity), while minimum-phase designs are shorter but introduce phase distortion

Multi-objective optimization and iterative design methods (like those of Johnston and Jain-Crochiere) help navigate these trade-offs for specific application requirements.

📡Advanced Signal Processing Unit 5 Review