Hubbry Logo
Sub-band codingSub-band codingMain
Open search
Sub-band coding
Community hub
Sub-band coding
logo
8 pages, 0 posts
0 subscribers
Be the first to start a discussion here.
Be the first to start a discussion here.
Sub-band coding
Sub-band coding
Sub-band coding
View on Wikipedia
from Wikipedia
Sub-band coding and decoding signal flow diagram

In signal processing, sub-band coding (SBC) is any form of transform coding that breaks a signal into a number of different frequency bands, typically by using a fast Fourier transform, and encodes each one independently. This decomposition is often the first step in data compression for audio and video signals.

SBC is the core technique used in many popular lossy audio compression algorithms including MP3.

Encoding audio signals

[edit]

The simplest way to digitally encode audio signals is pulse-code modulation (PCM), which is used on audio CDs, DAT recordings, and so on. Digitization transforms continuous signals into discrete ones by sampling a signal's amplitude at uniform intervals and rounding to the nearest value representable with the available number of bits. This process is fundamentally inexact, and involves two errors: discretization error, from sampling at intervals, and quantization error, from rounding.

The more bits used to represent each sample, the finer the granularity in the digital representation, and thus the smaller the quantization error. Such quantization errors may be thought of as a type of noise, because they are effectively the difference between the original source and its binary representation. With PCM, the audible effects of these errors can be mitigated with dither and by using enough bits to ensure that the noise is low enough to be masked either by the signal itself or by other sources of noise. A high quality signal is possible, but at the cost of a high bitrate (e.g., over 700 kbit/s for one channel of CD audio). In effect, many bits are wasted in encoding masked portions of the signal because PCM makes no assumptions about how the human ear hears.

Coding techniques reduce bitrate by exploiting known characteristics of the auditory system. A classic method is nonlinear PCM, such as the μ-law algorithm. Small signals are digitized with finer granularity than are large ones; the effect is to add noise that is proportional to the signal strength. Sun's Au file format for sound is a popular example of mu-law encoding. Using 8-bit mu-law encoding would cut the per-channel bitrate of CD audio down to about 350 kbit/s, half the standard rate. Because this simple method only minimally exploits masking effects, it produces results that are often audibly inferior compared to the original.

Basic principles

[edit]

The utility of SBC is perhaps best illustrated with a specific example. When used for audio compression, SBC exploits auditory masking in the auditory system. Human ears are normally sensitive to a wide range of frequencies, but when a sufficiently loud signal is present at one frequency, the ear will not hear weaker signals at nearby frequencies. We say that the louder signal masks the softer ones.

The basic idea of SBC is to enable a data reduction by discarding information about frequencies which are masked. The result differs from the original signal, but if the discarded information is chosen carefully, the difference will not be noticeable, or more importantly, objectionable.

First, a digital filter bank divides the input signal spectrum into some number (e.g., 32) of subbands. The psychoacoustic model looks at the energy in each of these subbands, as well as in the original signal, and computes masking thresholds using psychoacoustic information. Each of the subband samples are quantized and encoded so as to keep the quantization noise below the dynamically computed masking threshold. The final step is to format all these quantized samples into groups of data called frames, to facilitate eventual playback by a decoder.

Decoding is much easier than encoding, since no psychoacoustic model is involved. The frames are unpacked, subband samples are decoded, and a frequency-time mapping reconstructs an output audio signal.

Applications

[edit]

Beginning in the late 1980s, a standardization body, the Moving Picture Experts Group (MPEG), developed standards for coding of both audio and video. Subband coding resides at the heart of the popular MP3 format (more properly known as MPEG-1 Audio Layer III), for example.

Sub-band coding is used in the G.722 codec which uses sub-band adaptive differential pulse code modulation (SB-ADPCM) within a bit rate of 64 kbit/s. In the SB-ADPCM technique, the frequency band is split into two sub-bands (higher and lower) and the signals in each sub-band are encoded using ADPCM.

[edit]
Revisions and contributorsEdit on WikipediaRead on Wikipedia

Sub-band coding

View on Grokipedia
from Grokipedia
Sub-band coding is a technique that decomposes an input signal into multiple subbands using a bank of bandpass filters, followed by decimation to reduce the sampling rate in each subband, independent quantization and encoding of the subband signals with optimized bit allocation, and reconstruction via and synthesis filtering to approximate the original signal. This approach exploits the varying perceptual importance and statistical properties of different components to achieve efficient compression while minimizing , particularly in applications like audio and image processing. The core components of sub-band coding include analysis filter banks for signal decomposition, which apply low-pass and high-pass filters to separate frequency bands, and synthesis filter banks for reconstruction, often designed using quadrature mirror filters (QMFs) to ensure perfect reconstruction with minimal and phase distortion. Decimation by a factor of two in each subband halves the data rate per band, enabling adaptive bit allocation based on subband energy or perceptual models to prioritize bits for perceptually sensitive frequencies. Quantization introduces errors that are controlled across the frequency spectrum, and techniques like further reduce redundancy within subbands. Sub-band coding emerged in the 1970s as an extension of multirate signal processing, with foundational work on filter banks for perfect reconstruction by Croisier et al. in 1976 and detailed theoretical development in Crochiere and Rabiner's 1983 book Multirate Digital Signal Processing. It gained prominence in the 1980s for speech compression and evolved in the 1990s through connections to wavelet transforms, as explored in Vetterli and Kovačević's 1995 book Wavelets and Subband Coding, which unified subband methods with multiresolution analysis for better energy compaction. Advances in filter design and computational efficiency have made it a cornerstone of modern compression standards. Notable applications include audio coding in standards like MPEG-1 Audio Layer III () and Advanced Audio Coding (AAC), where subband decomposition enables perceptual noise shaping for high-fidelity compression at low bit rates. In image and video compression, it underpins , which uses wavelet-based subband coding for scalable, progressive transmission and superior performance over DCT-based at low bit rates. Other uses span for , hyperspectral image analysis, and error-resilient transmission in systems.

Fundamentals

Definition and Motivation

Sub-band coding (SBC) is a technique that decomposes an input signal into narrower sub-bands through the application of a bank of bandpass filters, permitting independent processing, quantization, and coding of each sub-band to facilitate efficient compression while targeting specific content. This decomposition exploits the -domain structure of the signal, allowing for reduced redundancy and tailored representation compared to time-domain methods. The motivation for sub-band coding arises from the limitations of uniform coding schemes like (PCM), which apply equal bit resolution across the entire signal spectrum and thus inefficiently allocate resources to high-frequency components that contribute little to perceptual quality. Instead, SBC leverages signal statistics—such as the concentration of energy in lower-frequency bands—and perceptual models, including auditory or visual masking where stronger signals obscure weaker ones in nearby frequencies, to enable non-uniform bit allocation that minimizes bitrate without audible or visible degradation. This approach is particularly advantageous for bandwidth-constrained applications like audio and transmission, where preserving subjective is paramount. In its basic model, an input discrete-time signal x(n)x(n) is divided into sub-bands via filter banks, with each sub-band signal decimated to lower its sampling rate, quantized based on perceptual relevance, and encoded for transmission or storage at a reduced overall bitrate. Reconstruction synthesizes the original signal by the coded sub-bands, applying synthesis filters, and summing the results to achieve at lower data rates. A representative example in audio compression illustrates this efficiency: (CD) audio uses 16-bit PCM at a stereo bitrate of 1.411 Mbit/s (44.1 kHz sampling), but early sub-band coders like the perceptual coder achieve near-CD quality at approximately 110 kbps per channel by assigning fewer bits to masked high frequencies and exploiting sub-band decimation.

Historical Development

Sub-band coding emerged in the 1970s as an application of multirate techniques to speech compression, building on foundational work in decimation and . Early explorations demonstrated that dividing speech signals into sub-bands allowed for more efficient quantization and bit allocation, reducing overall coding rates while controlling noise. A seminal contribution was the 1976 paper by Crochiere, Webber, and Flanagan, which proposed digitally coding speech in sub-bands using adaptive bit allocation based on perceptual importance, achieving significant bitrate reductions for applications. Concurrently, Crochiere and Rabiner advanced the theoretical underpinnings through their work on multirate processing, providing tools for efficient sub-band decomposition without excessive computational overhead. In the , advancements in propelled sub-band coding toward practical implementation, particularly through the development of quadrature mirror filters (QMFs) that enabled near-perfect reconstruction with minimal . At Bell Laboratories, J. D. Johnston introduced a family of optimized filters specifically tailored for QMF banks in 1980, improving frequency selectivity and reconstruction quality for audio signals. This era also saw the first international standardization with ITU-T G.722 in 1988, a sub-band (SB-ADPCM) operating at 64 kbit/s for (7 kHz), marking a milestone in high-quality speech transmission over digital networks. The 1990s integrated sub-band coding into multimedia standards, blending it with transform techniques for broader applications. The MPEG-1 Audio standard, finalized in 1992, incorporated a hybrid sub-band/transform filter bank in its Layer III () profile, enabling efficient compression of music at bitrates around 128 kbit/s per channel and revolutionizing distribution. Martin Vetterli and Jelena Kovačević's 1995 book Wavelets and Subband Coding formalized the deep connections between sub-band methods and wavelet theory, influencing subsequent designs by emphasizing multiresolution analysis for signal representation. Modern developments extended sub-band coding's efficiency into the late 1990s and beyond, with Advanced Audio Coding (AAC) standardized in 1997 as a perceptual coder using MDCT-based sub-bands for multichannel audio at lower bitrates than MP3. High-Efficiency AAC (HE-AAC), introduced in 2003, further enhanced low-bitrate performance through spectral band replication, achieving transparent quality at 48 kbit/s for stereo. In imaging, (2000) adopted wavelet sub-band decomposition for scalable compression, supporting lossless to lossy modes and outperforming DCT-based in visual quality.

Signal Decomposition and Analysis

Filter Banks

In sub-band coding, the analysis serves as the core mechanism for decomposing the input signal into multiple sub-bands, enabling efficient representation and subsequent processing. It consists of a parallel array of bandpass filters Hk(z)H_k(z), indexed by k=0k = 0 to M−1M-1, where MM denotes the number of sub-bands, each designed to isolate a specific portion of the signal's . Following , each sub-band signal undergoes downsampling by the factor MM, which reduces the sampling rate and data volume while preserving essential spectral content, thereby achieving critical sampling in filter banks. The output of the kk-th downsampler, representing the decimated sub-band signal, is expressed as yk(m)=∑nhk(n−mM)x(n),y_k(m) = \sum_{n} h_k(n - mM) x(n), where hk(n)h_k(n) is the of the kk-th analysis filter and x(n)x(n) is the input signal; this convolution-decimation operation efficiently extracts the sub-band components without redundant computation. A foundational type of filter bank is the two-channel quadrature mirror filter (QMF) bank, which performs critically sampled decomposition by splitting the signal into low- and high-frequency sub-bands using a and its mirror image, with cancellation properties inherent to the QMF . For broader applications requiring more sub-bands, polyphase implementations enhance efficiency by restructuring the into polyphase components followed by a computationally lightweight modulation stage, reducing the overall operation count in multirate systems. Design principles for these filter banks emphasize maximizing frequency selectivity to sharply delineate sub-bands, thereby minimizing energy leakage, while simultaneously suppressing introduced by downsampling through appropriate filter and transition band control. (FIR) filters are often preferred for their characteristics and stability, though (IIR) filters can offer sharper responses at lower computational cost in some configurations. From a computational perspective, multirate processing in filter banks eliminates redundancy by aligning decimation with filtration, achieving up to MM-fold reduction in processing load compared to non-multirate alternatives. Additionally, intentional overlap in the responses of adjacent filters ensures smooth transitions and avoids discontinuities at sub-band boundaries, supporting seamless signal .

Sub-band Division Process

The sub-band division process in sub-band coding constitutes the stage, where the input signal is decomposed into frequency-specific sub-bands to enable efficient processing and compression. This begins with filtering the input signal using an , which applies bandpass filters to isolate distinct components, such as lowpass and highpass filters denoted as H0(z)H_0(z) and H1(z)H_1(z) in the z-domain. Each sub-band signal is then downsampled by an integer factor, typically 2 or NN for an NN-channel bank, to reduce the sampling rate and data volume while preserving essential information; this decimation stretches the spectrum of each sub-band. Optionally, further transformations like the (DCT) may be applied within individual sub-bands to achieve additional decorrelation and energy compaction, particularly in image or audio applications. Frequency partitioning during this process can be uniform, dividing the signal spectrum into equal bandwidth sub-bands, or critical (non-uniform), such as octave-like bands that align with perceptual models like the Bark scale in audio to better match human psychoacoustics. Uniform partitioning is straightforward for fixed-rate systems, while critical partitioning optimizes for varying signal energy distribution across frequencies, often using logarithmic spacing to emphasize lower frequencies where perceptual sensitivity is higher. To mitigate introduced by downsampling, filters—typically lowpass filters with a cutoff at π/N\pi/N—are employed prior to decimation, ensuring that overlap from adjacent bands is minimized or canceled through careful . Quadrature mirror filters (QMF) serve as a common choice for this analysis stage due to their ability to provide near-perfect cancellation in critically sampled systems. A representative example is a 32-sub-band audio for signals from 0 to 20 kHz sampled at 48 kHz, where the spectrum is partitioned into uniform bands of approximately 625 Hz each to facilitate perceptual coding. The signal flow can be implemented via parallel filter banks, where all sub-bands are processed simultaneously, or tree-structured banks for multi-resolution analysis, involving iterative decomposition (e.g., applying lowpass/highpass pairs successively to create bands). In the parallel configuration, the input feeds into multiple analysis filters followed by downsamplers; tree structures, by contrast, cascade filters hierarchically, downsampling at each level to build a pyramid of resolutions suitable for progressive coding.

Coding and Quantization

Quantization Techniques

Quantization in sub-band coding involves mapping the continuous amplitude values of sub-band coefficients, obtained after signal decomposition, to a of discrete levels to achieve data rate reduction while minimizing perceptual . This introduces quantization , which must be controlled to remain inaudible or imperceptible. Two primary approaches are scalar quantization, applied independently to each coefficient, and , which processes groups of coefficients jointly to exploit statistical dependencies. Scalar quantization is widely used due to its low computational complexity and is typically either uniform, with equal step sizes across the range, or non-uniform, with varying step sizes to better match signal distributions or perceptual characteristics. In uniform scalar quantization, the coefficient yy is quantized as q=\round(y/Δ)⋅Δq = \round(y / \Delta) \cdot \Delta, where Δ\Delta is the fixed quantization step size and \round\round denotes rounding to the nearest integer. Non-uniform variants, such as companded quantization, apply a nonlinear mapping before uniform quantization to allocate finer resolution to smaller amplitudes, improving signal-to-noise ratio for low-level signals. Vector quantization offers higher efficiency by treating multiple sub-band coefficients as a vector and mapping it to the nearest codeword from a predefined , potentially capturing inter-coefficient correlations for better compression at low bit rates. However, its higher complexity limits widespread adoption in real-time sub-band coders, where scalar methods predominate. Perceptual quantization tailors the process to human sensory models, ensuring quantization noise falls below psychoacoustic masking thresholds to achieve transparency at reduced . In audio applications, bit allocation assigns fewer bits to sub-bands where signal energy is below the absolute hearing threshold or masked by stronger components, concentrating resources on perceptually salient regions. The step size Δ\Delta is often set inversely proportional to the masking level, such that higher masking allows coarser quantization without audible artifacts. Adaptive quantization dynamically adjusts the step size based on local signal energy or perceptual criteria within each sub-band, providing finer for high-energy segments and coarser for low-energy ones to optimize overall noise distribution. Noise shaping further refines this by spectral or temporal redistribution of quantization error, pushing it into or time regions of high masking to enhance perceived . In audio sub-band coding, simultaneous masking—arising from frequency proximity—and temporal masking—due to onset/offset effects—are modeled to determine sub-band-specific step sizes Δk\Delta_k for the kk-th sub-band, ensuring noise remains below the combined masking threshold. This approach, rooted in psychoacoustic principles, enables high-fidelity compression as demonstrated in standards like MPEG Audio.

Bit Allocation and Entropy Coding

In sub-band coding, bit allocation dynamically distributes a limited total bitrate across sub-bands to minimize reconstruction distortion while satisfying the rate constraint. This optimizes rate-distortion performance by assigning more bits to sub-bands with higher signal variance or perceptual significance, often through minimization of (MSE) under perceptual weighting. The approach ensures efficient use of bits by prioritizing sub-bands that contribute most to overall quality, treating the quantized sub-band outputs as inputs for this allocation. For signals modeled as parallel Gaussian sources, the water-filling algorithm provides an optimal strategy, iteratively pouring "water" (bits) into sub-bands to equalize the up to a common level, allocating zero bits to sub-bands below this threshold. This method, adapted from , maximizes the total rate or minimizes for a given rate by favoring stronger sub-bands. The optimal bit assignment for each sub-band kk is derived from rate-distortion theory as bk=12log⁡2(σk2λ),b_k = \frac{1}{2} \log_2 \left( \frac{\sigma_k^2}{\lambda} \right), where σk2\sigma_k^2 denotes the variance of the sub-band signal and λ\lambda is the Lagrange multiplier adjusted to meet the total bitrate constraint; integer rounding and iterative refinement are applied in practice for feasibility. In perceptual applications like audio coding, bit allocation incorporates psychoacoustic models to weight sub-bands based on human auditory masking thresholds, ensuring imperceptible distortion. These models, as specified in ISO MPEG guidelines, compute signal-to-mask ratios to guide allocation, emphasizing tonal and noise-like components while de-emphasizing masked regions. For instance, in the MP3 audio standard, variable bit allocation is applied across 576 frequency-domain samples per granule, dynamically adjusting bits per sub-band to balance compression and perceptual fidelity. Post-quantization, entropy coding compresses the allocated bitstream by exploiting symbol probabilities, further reducing redundancy without loss. assigns variable-length prefix codes to quantized indices, with shorter codes for frequent values like near-zero coefficients, achieving near- rates for typical sub-band distributions. offers superior efficiency by encoding entire sequences into a single fractional number, approaching the theoretical limit more closely than Huffman, especially for non-integer bit requirements. In sparse high-frequency sub-bands, where many coefficients are zero, complements these methods by efficiently representing consecutive zeros as (value, length) pairs, minimizing bits for insignificant details.

Reconstruction and Synthesis

Synthesis Filter Banks

In sub-band coding, the synthesis filter bank reconstructs the original signal from the quantized and coded sub-band signals by first each sub-band component and then applying low-pass or band-pass filtering followed by summation across all bands. Specifically, each quantized sub-band signal yk(m)y_k(m), for k=0,1,…,M−1k = 0, 1, \dots, M-1, is upsampled by the decimation factor MM through zero-insertion, which expands the sampling rate and replicates the , after which it is filtered by the synthesis filter gk(n)g_k(n) (or Gk(z)G_k(z) in the z-domain) to interpolate and suppress artifacts before the outputs are added to form the reconstructed signal x^(n)\hat{x}(n). This inverts the stage, where the input signal was decomposed into sub-bands via downsampling and filtering. The design of synthesis filters typically employs mirror-image relationships to the corresponding filters in the counterpart bank, such that gk(n)=hk(−n)g_k(n) = h_k(-n), ensuring compatibility and facilitating management while maintaining properties in () implementations. For computational efficiency, polyphase representations decompose the synthesis filters into parallel branches, avoiding redundant operations in the upsampling and filtering cascade and reducing the overall structure to a in the polyphase domain. These designs are particularly effective in maximally decimated filter banks, where the number of channels equals the decimation factor MM, allowing the synthesis process to operate at the original sampling rate after combination. Aliasing distortions introduced during the downsampling in the stage are canceled in the synthesis bank through the coordinated filter responses, where the synthesis filters are crafted to nullify the shifted components arising from modulation effects in the sub-bands. This cancellation relies on the filters' frequency selectivity, ensuring that unwanted aliases from adjacent bands do not propagate into the reconstructed output. The reconstructed signal can be expressed in the time domain as x^(n)=∑k=0M−1∑m=−∞∞gk(n−mM) yk(m),\hat{x}(n) = \sum_{k=0}^{M-1} \sum_{m=-\infty}^{\infty} g_k(n - mM) \, y_k(m), where gk(n)g_k(n) denotes the of the kk-th synthesis filter, and the inner sum accounts for the upsampled nature of yk(m)y_k(m) by spacing the contributions every MM samples. This convolution-based formulation highlights the role of the synthesis filters in recovering the full-bandwidth signal. In practical implementations, the of the synthesis is mitigated through polyphase structures, which achieve linear scaling with filter length, but FFT-based methods for modulated or DFT further reduce it to O(Mlog⁡M)O(M \log M) operations per output sample, making them suitable for real-time audio and applications despite the multi-channel .

Conditions for Perfect Reconstruction

Perfect reconstruction (PR) in sub-band coding refers to the ability of a to recover the original input signal X(z)X(z) exactly as a delayed version X^(z)=z−lX(z)\hat{X}(z) = z^{-l} X(z), where ll is an delay, without any distortion or artifacts. This property is essential for lossless decomposition and reconstruction, ensuring that the coding process does not introduce irreversible errors in the absence of quantization. In a two-channel quadrature mirror filter (QMF) , PR is achieved when the filters H0(z)H_0(z) (low-pass) and H1(z)H_1(z) (high-pass), along with synthesis filters G0(z)G_0(z) and G1(z)G_1(z), satisfy specific conditions derived from the polyphase representation. The function must equal H0(z)G0(z)+H1(z)G1(z)=2z−lH_0(z) G_0(z) + H_1(z) G_1(z) = 2z^{-l}, ensuring no or phase beyond the delay, while the cancellation condition requires H1(z)G0(z)−H0(z)G1(z)=0H_1(z) G_0(z) - H_0(z) G_1(z) = 0. These equations arise from the overall of the critically sampled system (decimation factor of 2), where terms from downsampling are eliminated by appropriate choice of synthesis filters, such as G0(z)=H1(−z)G_0(z) = H_1(-z) and G1(z)=−H0(−z)G_1(z) = -H_0(-z). For advanced designs, paraunitary filter banks enable orthogonal PR, where the and synthesis filters form an , preserving energy and allowing efficient implementation via lattice structures. In wavelet filter banks, PR is extended to multi-resolution , supporting hierarchical sub-band decomposition with compactly supported filters that maintain across scales. A key states that PR is possible if the filters are biorthogonal, meaning the and synthesis filter sets satisfy inner product conditions that ensure invertibility without . In practice, quantization during coding introduces errors that prevent exact PR, leading to imperfect reconstruction where the error signal ee degrades the output. This degradation is quantified by the (SNR), defined as SNR=10log⁡10(Px/Pe)\text{SNR} = 10 \log_{10} (P_x / P_e), where PxP_x is the power of the original signal and PeP_e is the power of the quantization error; higher SNR values indicate better , typically targeted above 30 dB in audio applications. Oversampled filter banks, where the decimation factor is less than the number of channels (introducing redundancy), relax the PR conditions by providing frame expansions that allow robust reconstruction even with imperfect filters, as the additional samples compensate for aliasing and distortion.

Applications

Audio Compression

Sub-band coding plays a pivotal role in audio compression by decomposing the signal into frequency subbands, allowing efficient encoding that exploits human auditory perception within the typical hearing range of 20 Hz to 20 kHz. This approach enables significant bitrate reduction while maintaining perceptual quality, as it discards inaudible components and allocates bits based on psychoacoustic models that account for masking effects. In early standards like Layer I, audio-specific adaptations employ a 32-subband quadrature mirror filter (QMF) bank to divide the input signal into equal-width bands, supporting sampling rates of 32, 44.1, and 48 kHz at bitrates from 32 to 448 kbit/s. This subband structure facilitates uniform quantization across bands, with performance demonstrated in reducing the uncompressed CD audio bitrate of 1411 kbit/s (for 16-bit, 44.1 kHz stereo) to levels like 192 kbit/s while preserving near-transparent quality for most listeners. A notable example is the format ( Layer III), which uses a hybrid combining a polyphase QMF for initial 32-subband division with a (MDCT) to further refine resolution, achieving 128 kbit/s for stereo audio; here, the polyphase splits the signal into 32 bands, which are then quantized according to a psychoacoustic model that determines masking thresholds for bit allocation. Other codecs illustrate sub-band coding's versatility in audio applications. The G.722 standard implements sub-band (SB-ADPCM), splitting the signal into two bands to cover a 7 kHz bandwidth at 64 kbit/s, providing speech quality superior to alternatives. (AAC), an evolution with roots in sub-band techniques, primarily uses a 1024-point MDCT but incorporates perceptual subband-like processing to support sampling rates up to 96 kHz across multiple channels. In modern usage, the Opus , standardized in 2012, employs a hybrid approach integrating sub-band coding elements with and MDCT for low-latency VoIP, enabling bitrates from 6 to 510 kbit/s with robust performance in real-time communication.

Image and Video Coding

Sub-band coding has been adapted for by extending one-dimensional filter banks to two-dimensional separable structures, where filters are applied separately along rows and columns to decompose images into frequency sub-bands. This approach enables multi-resolution analysis, capturing both low-frequency approximations and high-frequency details essential for visual fidelity. In , the standard utilizes the (DWT) implemented via such 2D filter banks, achieving efficient coding through sub-band decomposition into as many as 16 or more levels. For in , the reversible 5/3 LeGall filter is employed, featuring coefficients that support exact reconstruction without information loss. In contrast, the irreversible 9/7 is used for , offering superior energy compaction for high-frequency sub-bands at compression ratios such as 20:1 to 50:1, particularly effective for complex textures. These facilitate progressive transmission, where low-frequency sub-bands are prioritized for quick previews. Bit allocation in these systems can incorporate visual masking models, akin to psychoacoustic principles in audio, to allocate fewer bits to perceptually less sensitive sub-bands. Key processes in image sub-band coding include the Laplacian pyramid, which generates a series of difference images between Gaussian-smoothed versions at progressively reduced resolutions, forming bandpass sub-bands that enhance edge preservation. Lapped transforms, such as the modulated lapped transform (MLT), overlap adjacent blocks to reduce blocking artifacts, improving coding efficiency in non-stationary image regions. For encoding, embedded zerotree (EZW) coding exploits the statistical dependency of wavelet coefficients across sub-bands, treating insignificant coefficients as zerotrees rooted in parent-child relationships for scalable, embedded bitstreams suitable for progressive transmission. In performance, demonstrates superior compression to DCT-based , especially at high ratios above 20:1, where sub-bands preserve more details and reduce artifacts like ringing, achieving typically 1-3 dB higher PSNR in natural images. For video coding, sub-band techniques extend to three-dimensional decompositions, incorporating temporal dimensions alongside spatial ones to handle motion efficiently. Motion-compensated sub-band coding (SBC) appears in intra-frame processing within scalable extensions like H.264/AVC's SVC, where sub-bands are formed for enhancement layers using -like decompositions on residual frames. SVC employs multi-band layers for spatial, temporal, and quality scalability, allowing layered bitstreams that adapt to varying network conditions. A prominent example is the use of 3D wavelets with motion-compensated temporal filtering (MCTF), which applies temporal sub-bands across frames after motion compensation to decorrelate sequences, followed by spatial 2D wavelets on the resulting sub-bands for compact representation. This enables efficient handling of dynamic scenes, with temporal low-pass sub-bands capturing global motion and high-pass sub-bands isolating local changes. Video SBC via these methods supports SNR scalability, where finer quantization in higher sub-bands allows graceful quality degradation by truncating enhancement layers without disrupting base-layer decoding.

Other Applications

Sub-band coding is also applied in hyperspectral image analysis for efficient compression of multi-spectral data, using 3D decompositions to exploit spectral and spatial correlations while preserving fine spectral details essential for and material identification. Additionally, it enhances error-resilient transmission in systems by enabling unequal error protection, where more bits or redundancy are allocated to perceptually important low-frequency subbands, improving robustness against in networks.

Advantages and Limitations

Key Benefits

Sub-band coding provides significant efficiency gains by enabling adaptive coding tailored to the characteristics of individual frequency bands, exploiting the non-stationary nature of signals to allocate bits more effectively than uniform methods like (PCM). This approach can achieve bitrate reductions of up to 50%, such as compressing speech from 128 kbps to 64 kbps while maintaining quality, due to the coding gain derived from variance disparities across sub-bands. A key perceptual advantage lies in confining quantization errors to specific sub-bands where they are masked by the human auditory or visual system, ensuring distortions remain imperceptible and preserving transparency even at low bitrates. For instance, in audio compression, this psychoacoustic exploitation allows systems like MPEG-1 Audio Layer I—based on sub-band filtering—to deliver near-CD quality at 384 kbps, compared to the 1.411 Mbps required for uncompressed PCM, representing roughly a 1/3 reduction, with further layers achieving up to 1/10th the bitrate. The flexibility of sub-band coding supports scalable decoding from coarse to fine resolutions, facilitating progressive transmission and compatible with multi-resolution analysis through implementations that decompose signals hierarchically. Additionally, the independence of sub-bands enables parallel processing, allowing distributed computation across bands to reduce latency and computational load in real-time systems via pipelined or polyphase structures. In image compression, sub-band methods often outperform block-based transforms like DCT by providing better energy compaction and higher (PSNR), with less blocking artifacts at low rates, as demonstrated in comparative studies. These benefits manifest across applications, such as audio standards achieving perceptual transparency and image coders yielding superior reconstruction metrics.

Challenges and Drawbacks

Sub-band coding systems exhibit high primarily due to the implementation of and synthesis filter banks, which often require O(N log N) operations per frame in FFT-based designs for signal decomposition and reconstruction. This overhead can be mitigated through efficient algorithms such as the (MDCT), which reduces processing demands while maintaining critical sampling properties in audio applications. Imperfections in reconstruction lead to various artifacts, including ringing effects around edges, exacerbated by long synthesis filters and quantization noise. In critically sampled filter banks, aliasing artifacts arise from subsampling, manifesting as spectral overlaps that degrade signal quality unless compensated by perfect reconstruction conditions. These issues are particularly pronounced in image and video coding, where nonlinear phase responses introduce additional waveform distortions. Bit allocation in sub-band coding necessitates side information transmission, such as scale factors and quantization indices, which can constitute up to 14 kb/s per channel and represent a notable portion of the total bitrate (approximately 5-10% in mid-range scenarios). This overhead renders the system sensitive to channel errors during transmission, as corruption of side information can propagate distortions across subbands. At very low bitrates, sub-band coding proves less effective without hybrid approaches, such as combining it with or sinusoidal modeling, as energy compaction falters in high-frequency subbands leading to increased . involves inherent trade-offs between length and delay; longer filters improve frequency selectivity and reduce but introduce greater algorithmic delay (proportional to half the filter length in implementations). In the 2020s, traditional sub-band coding has been largely surpassed by neural audio codecs, which achieve superior perceptual quality at ultra-low bitrates through end-to-end learning, outperforming classical methods in efficiency and artifact suppression.

References

  1. https://www.sciencedirect.com/topics/computer-science/subband-coding
  2. https://persci.mit.edu/pub_pdfs/simoncelli_subband.pdf
  3. https://personal.utdallas.edu/~aria/papers/chapter.pdf
  4. http://www.mp3-tech.org/programmer/docs/audiopaper1.pdf
  5. https://lame.sourceforge.io/doc/vol5num2art3.pdf
  6. https://onlinelibrary.wiley.com/doi/abs/10.1002/j.1538-7305.1976.tb02929.x
  7. http://waveletsandsubbandcoding.org/Repository/VetterliKovacevic95_Manuscript.pdf
  8. https://jpeg.org/jpeg2000/
  9. https://link.springer.com/chapter/10.1007/978-1-4757-2119-5_2
  10. https://www.audiolabs-erlangen.de/content/resources/aesCodingTutorial/media/tutorial/aesbosi.pdf
  11. https://www.audiolabs-erlangen.de/content/resources/aesCodingTutorial/media/tutorial/tutorial.pdf
  12. https://www.sciencedirect.com/topics/computer-science/coded-quantization
  13. https://www.cns.nyu.edu/~david/courses/perceptionGrad/Readings/PainterSpanias-ProcIEEE2000.pdf
  14. https://jontalle.web.engr.illinois.edu/uploads/537.F18/Papers/BrandenburgFallerHerreJohnsontKleijn.13.pdf
  15. https://ieeexplore.ieee.org/document/871488/
  16. https://iphome.hhi.de/schwarz/assets/dc/VBpart1.pdf
  17. https://www.ee.columbia.edu/~dpwe/papers/Pan95-mpeg-ieeemm.pdf
  18. http://www.mp3-tech.org/programmer/docs/mp3_theory.pdf
  19. https://www.cambridge.org/core/books/digital-signal-compression/entropy-coding-techniques/D37695D95F81472E2A31595CBBD61E19
  20. https://www.utdallas.edu/~aria/papers/chapter.pdf
  21. https://engineering.purdue.edu/~ee538/QMF.pdf
  22. https://www.sciencedirect.com/topics/engineering/perfect-reconstruction-filter-banks
  23. https://ieeexplore.ieee.org/document/17560
  24. https://ieeexplore.ieee.org/document/106862
  25. https://web.stanford.edu/class/ee368b/Resources/girod:95-SubbandImageCoding.pdf
  26. https://ieeexplore.ieee.org/document/668788
  27. https://www.dspguide.com/ch22/1.htm
  28. https://mpeg.chiariglione.org/standards/mpeg-1/audio.html
  29. https://www.ee.columbia.edu/~dpwe/papers/Brand99-mp3.pdf
  30. https://web.stanford.edu/class/ee398a/handouts/lectures/10-JPEG2000.pdf
  31. http://helmut.knaust.info/class/201110_5311/papers/SVK.pdf
  32. https://engineering.purdue.edu/~ee538/Image_compression_wavelets_jpeg2000.pdf
  33. https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/Malvar_MLT_TSP0690.pdf
  34. https://www.asprs.org/wp-content/uploads/pers/2005journal/feb/2005_feb_157-167.pdf
  35. https://www.spiedigitallibrary.org/conference-proceedings-of-spie/6822/1/A-new-subband-wavelet-framework-for-AVC-H264-intraframe-coding/10.1117/12.767081.full
  36. https://www.lirmm.fr/~wpuech/recherche/publications/Book/11_Scalable-video-coding.pdf
  37. https://link.springer.com/chapter/10.1007/11596981_126
  38. https://www.spiedigitallibrary.org/conference-proceedings-of-spie/5150/1/Scalable-video-compression-using-longer-motion-compensated-temporal-filters/10.1117/12.502725.full
  39. https://projects.cwi.nl/mascot/papers/Bottreau_etal_icip01.pdf
  40. https://sites.ecse.rpi.edu/~pearlman/papers/tgrs04_tpm.pdf
  41. https://www.stanford.edu/class/ee368b/Resources/girod:95-SubbandImageCoding.pdf
  42. https://www.sciencedirect.com/science/article/pii/B9780124158931000135
  43. http://www.cs.umsl.edu/~sanjiv/classes/cs6420/lectures/wavelet.pdf
  44. https://www.itu.int/wftp3/av-arch/avc-site/1990-1996/Contributions/AVC-0149.pdf
  45. https://www.researchgate.net/publication/4333527_Subband_conversion_for_feature_extraction_from_compressed_audio
  46. http://www.mp3-tech.org/programmer/docs/jacaba_main.pdf
  47. https://www.mp3-tech.org/programmer/docs/dts_whitepaper.pdf
  48. https://www.mmsp.ece.mcgill.ca/Theses/1982/RabipourT1982.pdf
  49. https://ieeexplore.ieee.org/document/818217/
  50. https://www.onsemi.com/pub/Collateral/AND8382-D.PDF
  51. https://arxiv.org/abs/2510.15364
Add your contribution
Related Hubs
User Avatar
No comments yet.