Filter Bank Structure Research Articles

A Structurally Regularized CNN Architecture via Adaptive Subband Decomposition.

We propose a generalized convolutional neural network (CNN) architecture that first decomposes the input signal into subbands by an adaptive filter bank structure, and then uses convolutional layers to extract features from each subband independently. Fully connected (FC) layers finally combine the extracted features to perform classification. The proposed architecture restrains each of the subband CNNs from learning using the entire input signal spectrum, resulting in structural regularization. Our proposed CNN architecture is fully compatible with the end-to-end learning mechanism of typical CNN architectures and learns the subband decomposition from the input dataset. We show that the proposed CNN architecture has attractive properties, such as robustness to input and weight-and-bias quantization noise, compared to regular full-band CNN architectures. Importantly, the proposed architecture significantly reduces computational costs, while maintaining state-of-the-art classification accuracy. Experiments on image classification tasks using the MNIST, CIFAR-10/100, Caltech-101, and ImageNet-2012 datasets show that the proposed architecture allows accuracy surpassing state-of-the-art results. On the ImageNet-2012 dataset, we achieved top-5 and top-1 validation set accuracy of 86.91% and 69.73%, respectively. Notably, the proposed architecture offers over 90% reduction in computation cost in the inference path and approximately 75% reduction in back-propagation (per iteration) with just a single-layer subband decomposition. With a two-layer subband decomposition, the computational gains are even more significant with comparable accuracy results to the single-layer decomposition.

A position-adaptive noise-reduction method using a deep denoising filter bank for dedicated breast positron emission tomography images.

Dedicated breast positron emission tomography (db-PET) is more sensitive than whole-body positron emission tomography and is thus expected to detect early stage breast cancer and determine treatment efficacy. However, it is challenging to decrease the sensitivity of the chest wall side at the edge of the detector, resulting in a relative increase in noise and a decrease in detectability. Longer acquisition times and injection of larger amounts of tracer improve image quality but increase the burden on the patient. Therefore, this study aimed to improve image quality via reconstruction with shorter acquisition time data using deep learning, which has recently been widely used as a noise reduction technique. In our proposed method, a multi-adaptive denoising filter bank structure was introduced by training the training data separately for each detector area because the noise characteristics of db-PET images vary at different locations. Input and ideal images were reconstructed based on 1- and 7-min collection data, respectively, using list mode data. The deep learning model used residual learning with an encoder-decoder structure. The image quality of the proposed method was superior to that of existing noise reduction filters such as Gaussian filters and nonlocal mean filters. Furthermore, there was no significant difference between the maximum standardized uptake values before and after filtering using the proposed method. Taken together, the proposed method is useful as a noise reduction filter for db-PET images, as it can reduce the patient burden, scan time, and radiotracer amount in db-PET examinations.

General Closed-Form Transfer Function Expressions for Fast Filter Bank

The existing literature focuses on the applications of fast filter bank due to its excellent frequency responses with low complexity. However, the topic is not addressed related to the general transfer function expressions of the corresponding subfilters for a specific channel. To do this, in this paper, general closed-form transfer function expressions for fast filter bank are derived. Firstly, the cascaded structure of fast filter bank is modelled by a binary tree, with which the index of the subfilter at each stage within the channel can be determined. Then the transfer functions for the two outputs of a subfilter are expressed in a unified form. Finally, the general closed-form transfer functions for the channel and its corresponding subfilters are obtained by variables replacement if the prototype lowpass filters for the stages are given. Analytical results and simulations verify the general expressions. With such closed-form expressions lend themselves easily to analysis and direct computation of the transfer functions and the frequency responses without the structure graph.

Vibration-based looseness identification of bolted structures via quasi-analytic wavelet packet and optimized large margin distribution machine

Bolted joints are the most widely utilized connection types in industries, and therein looseness identification of bolted structures is of great significance to guarantee structural reliability. In this article, a comprehensive study of bolt looseness identification under random excitation is presented. To fulfill this task, this research focuses on three prominent difficulties, including nonstationary signal processing, subtle feature extraction, and robust state classification. First, a novel filter bank structure of quasi-analytic dual-tree complex wavelet packet transform is constructed to analyze the measured vibration response signals, for purpose of capturing subtle feature information. Then, multiple features are extracted from subband signals to capture the variations of dynamic characteristics, and sensitive features are selected by Laplacian score to construct the low-dimensional feature set. Subsequently, a novel classifier with better generalization performance, named large margin distribution machine, is optimized with the wavelet kernel function and the whale optimization algorithm, in order to handle the intrinsic uncertainty related to the looseness states of bolted structures. After feeding the low-dimensional feature set, the proposed classifier is trained to identify looseness states of bolted structures. Finally, experiments of a two-bolt lapped beam under random excitation are conducted to verify the effectiveness of the proposed method, and two typical loading conditions (paired-bolt looseness and single-bolt looseness) are considered. Besides, the superiority of the proposed method is demonstrated by comparing with other analogical methods. This research can provide a promising implement in practical applications of bolt looseness identification under random excitation.

Filter bank sinc-convolutional network with channel self-attention for high performance motor imagery decoding

Objective. Motor Imagery Brain-Computer Interface (MI-BCI) is an active Brain-Computer Interface (BCI) paradigm focusing on the identification of motor intention, which is one of the most important non-invasive BCI paradigms. In MI-BCI studies, deep learning-based methods (especially lightweight networks) have attracted more attention in recent years, but the decoding performance still needs further improving. Approach. To solve this problem, we designed a filter bank structure with sinc-convolutional layers for spatio-temporal feature extraction of MI-electroencephalography in four motor rhythms. The Channel Self-Attention method was introduced for feature selection based on both global and local information, so as to build a model called Filter Bank Sinc-convolutional Network with Channel Self-Attention for high performance MI-decoding. Also, we proposed a data augmentation method based on multivariate empirical mode decomposition to improve the generalization capability of the model. Main results. We performed an intra-subject evaluation experiment on unseen data of three open MI datasets. The proposed method achieved mean accuracy of 78.20% (4-class scenario) on BCI Competition IV IIa, 87.34% (2-class scenario) on BCI Competition IV IIb, and 72.03% (2-class scenario) on Open Brain Machine Interface (OpenBMI) dataset, which are significantly higher than those of compared deep learning-based methods by at least 3.05% ( = 0.0469), 3.18% ( = 0.0371), and 2.27% ( = 0.0024) respectively. Significance. This work provides a new option for deep learning-based MI decoding, which can be employed for building BCI systems for motor rehabilitation.

Per-Wavelet Equalization for Discrete Wavelet Transform Based Multi-Carrier Modulation Systems

The Discrete Wavelet Transform (DWT) has gained attention in the area of Multi-Carrier Modulation (MCM) because it can overcome some well known limitations of Discrete Fourier Transform (DFT) based MCM systems. Its improved spectral containment removes the need for a cyclic prefix, be it that appropriate equalization then has to be added as the cyclic convolution property no longer holds. Most DWT based MCM systems in the literature use Time-domain EQualizers (TEQs) to mitigate the channel distortion. In this paper, a Per-Wavelet EQualizer (PWEQ) is proposed which directly maximizes the Signal-to-Interference-plus-Noise Ratio (SINR) per symbol and is applicable to any wavelet family. The proposed PWEQ provides a performance upper bound for the TEQs for DWT based MCM systems. The computational complexity of the PWEQ is reduced by modifying the Filter Bank (FB) structure of the DWT. Simulations are performed to compare the PWEQ performance against the TEQs for DWT based MCM systems and the similar Per-Tone EQualizer (PTEQ) for DFT based MCM systems. The simulations are performed using measured Asymmetric Digital Subscriber Line (ADSL) and G.fast channels with Fejér-Korovkin (FK) wavelets. The proposed PWEQ increases the SINR on the received symbols compared to the TEQs at the cost of an increased computational complexity.

Least-Squares Equalizer Demonstrations Using a Full-Digital Bandwidth Sub-Nyquist-Sampled Wideband Beamformer on an RFSoC

The radio frequency system-on-a-chip (RFSoC) has recently become a viable candidate for completely replacing traditional analog and digital front ends, facilitating the development of wideband phased-array systems where the modern-day RFSoC takes the comprehensive, dominate role in the architecture of the array. Wideband phased-array systems require high-fidelity compensation techniques capable of correcting imbalanced and dispersive channel effects for effective beamforming. This article provides solutions to these challenges by designing a wideband equalizer for a sub-Nyquist-sampled 1.6-GHz <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">S</i> -band phased-array system implemented on a Xilinx 8-channel RFSoC, whose analog-to-digital converters (ADC) operate at 4 gigasamples per second. In brief, an RFSoC is a unique, state-of-the-art, highly integrated device that incorporates a field programmable gate array, high-speed ADCs and digital-to-analog converters with a system-on-a-chip architecture on a single monolithic device. By definition, true time delay beamsteering can be implemented digitally via a combination of integer-sample delays and fractional-sample delay finite impulse response filters. By modifying the filter structure of the fractional-sample delay filter bank to support complex coefficients, channel equalization is integrated with fractional-sample delay filters to mitigate undesired channel effects. For the first time, to the best of our knowledge, this article has developed an equalizer design methodology for an uncalibrated 8-element RFSoC-based sub-Nyquist-sampled wideband beamformer. Lab measurements confirm efficacy.

Instantaneous Frequency Selective Filtering Using Ensemble Empirical Mode Decomposition

In this paper, novel instantaneous frequency (IF) selective high-pass, low-pass and band-pass filtering techniques for multicomponent signals are developed using the ensemble empirical mode decomposition algorithm (EEMD). The EEMD algorithm is based on the property that the empirical mode decomposition algorithm acts on fractional Gaussian noise as a dyadic filter bank of constant-Q band-pass filters. Unlike pre-determined sub-band filtering, the filter-bank structure observed for EEMD applies locally to a signal. In the proposed techniques, frequency translation is carried out to ensure that the ratio of edge IFs of the pass-band and the stop-band is such that the EEMD algorithm will be able to distinguish between the components in the two bands. Also, EEMD approach is applied with band-limited white noise so that the signal components in the pass band and the stop band are extracted in different intrinsic mode functions. Simulations are used to demonstrate the efficacy of the proposed filtering algorithms and to compare their performance with conventional filtering techniques. The performance of the filters is assessed subjectively and in terms of objective criteria in presence of noise. The proposed filtering technique is also applied on a real speech signal to isolate speech resonance signals in accordance with the AM-FM model of speech.

Traversal index enhanced-gram (TIEgram): A novel optimal demodulation frequency band selection method for rolling bearing fault diagnosis under non-stationary operating conditions

It is very important to select the optimal demodulation frequency band (ODFB) of rolling bearing vibration signals for the most valuable fault information extraction and diagnosis. Fast kurtogram (FK) is an effective and most commonly used ODFB selection approach for bearing fault diagnosis, which generally is founded on the filter bank structure and short-time Fourier transform. Though the FK method can effectively detect the shock characteristics of frequency band signals, other useful characteristics related with failure of vibration signal will be ignored. In this paper, a novel ODFB selection method called traversal index enhanced-gram (TIEgram) is proposed for rolling bearing vibration signals. In the proposed TIEgram method, first of all, the traversal segmentation model is utilized to transfer the original signal into frequency domain for enhancing overall segmentation performance of traditional binary trees and 1/3 binary trees structure segmentation models. Then, a new weighted fusion indicator based on the kurtosis, correlation coefficient and spectral negative entropy is designed to select ODFB from the segmented results of traversal segmentation model, which can effectively solve the problem that different vibration signal characteristics cannot be fully detected by a single indicator. After that, an enhanced adaptive multi-scale weighted morphological filtering-based envelope spectrum is employed to demodulate the obtained frequency band for a much more accurate diagnosis effect of rolling bearing. Finally, the simulated and measured signals of rolling bearing under stationary and non-stationary operating conditions are respectively used to verify the feasibility and effectiveness of the proposed method with comparison of the existing FK, Autogram and infogram methods. The comparison analysis results show that TIEgram method can accurately identify the most useful fault information and shows better performance than existing methods.

A Modified Complex Variational Mode Decomposition Method for Analyzing Nonstationary Signals with the Low-Frequency Trend

Complex variational mode decomposition (CVMD) has been proposed to extend the original variational mode decomposition (VMD) algorithm to analyze complex-valued data. Conventionally, CVMD divides complex-valued data into positive and negative frequency components using bandpass filters, which leads to difficulties in decomposing signals with the low-frequency trend. Moreover, both decomposition number parameters of positive and negative frequency components are required as prior knowledge in CVMD, which is difficult to satisfy in practice. This paper proposes a modified complex variational mode decomposition (MCVMD) method. First, the complex-valued data are upsampled through zero padding in the frequency domain. Second, the negative frequency component of upsampled data are shifted to be positive. Properties of analytical signals are used to get the real-valued data for standard variational mode decomposition and the complex-valued decomposition results after frequency shifting back. Compared with the conventional method, the MCVMD method gives a better decomposition of the low-frequency signal and requires less prior knowledge about the decomposition number. The equivalent filter bank structure is illustrated to analyze the behavior of MCVMD, and the MCVMD bi-directional Hilbert spectrum is provided to give the time–frequency representation. The effectiveness of the proposed algorithm is verified by both synthetic and real-world complex-valued signals.

FB-CGANet: filter bank channel group attention network for multi-class motor imagery classification

Objective. Motor imagery-based brain–computer interface (MI-BCI) is one of the most important BCI paradigms and can identify the target limb of subjects from the feature of MI-based Electroencephalography signals. Deep learning methods, especially lightweight neural networks, provide an efficient technique for MI decoding, but the performance of lightweight neural networks is still limited and need further improving. This paper aimed to design a novel lightweight neural network for improving the performance of multi-class MI decoding. Approach. A hybrid filter bank structure that can extract information in both time and frequency domain was proposed and combined with a novel channel attention method channel group attention (CGA) to build a lightweight neural network Filter Bank CGA Network (FB-CGANet). Accompanied with FB-CGANet, the band exchange data augmentation method was proposed to generate training data for networks with filter bank structure. Main results. The proposed method can achieve higher 4-class average accuracy (79.4%) than compared methods on the BCI Competition IV IIa dataset in the experiment on the unseen evaluation data. Also, higher average accuracy (93.5%) than compared methods can be obtained in the cross-validation experiment. Significance. This work implies the effectiveness of channel attention and filter bank structure in lightweight neural networks and provides a novel option for multi-class motor imagery classification.

Infrared and visible image fusion based on QNSCT and Guided Filter

Image fusion is the process of fusing multiple images of the same scene to obtain a more informative image for human eye perception. In this paper, a new fusion framework based on Quaternion Non-Subsampled Contourlet Transform (QNSCT) and Guided Filter detail enhancement is designed to address the problems of inconspicuous infrared targets and poor background texture in Infrared and visible image fusion. The proposed method uses the quaternion wavelet transform for the first time instead of the traditional Non-Subsampled Pyramid Filter Bank structure in the Non-Subsampled Contourlet Transform (NSCT). The flexible multi-resolution of quaternion wavelet and the multi-directionality of NSCT are fully utilized to refine the multi-scale decomposition scheme. On the other hand, the coefficient matrix obtained from the proposed QNSCT algorithm is fused using a weight refinement algorithm based on the guided filter. The fusion scheme is divided into four steps. First, the Infrared and visible images are decomposed into multi-directional and multiscale coefficient matrices using QNSCT. The experimental results show that the proposed algorithm not only extracts important visual information from the source image, but also preserves the texture information in the scene better. Meanwhile, the scheme outperforms state-of-the-art methods in both subjective and objective evaluations.

Reconfigurable Nonuniform Filter Bank for Hearing Aid Systems

Filter banks with good reconfigurability are desired in hearing aid systems due to the individual requirements of patients with different hearing loss characteristics. This paper proposes a modularized design approach of a completely reconfigurable nonuniform filter bank for hearing aid systems. The proposed filter bank structure consists of a multiband generation module and a subband extraction module, where the frequency warping technique is adopted. Based on theoretical analyses of second order frequency warping, the subband distribution can be flexibly controlled. Moreover, to adapt the subband extraction module design to the multiband generation module, the relationship between control parameters is analyzed and a mapping formula is derived. Through the co-design of the two modules, superior reconfigurability over existing methods is achieved. Application examples show that the proposed filter bank is able to provide multiple subbands distribution schemes that satisfy the requirements of different audiograms. The proposed filter bank structure has lower complexity, smaller delay and smaller matching errors than existing filter bank structures due to its flexible structure.

Filter bank riemannian-based kernel support vector machine for motor imagery decoding

Brain computer interface (BCI) enables the communication between the brain and external machines through Electroencephalography (EEG) signals, which has attracted lots of attention. Motor Imagery-based BCI (MI-BCI) is one of the most important paradigms in the BCI field. In MI-BCI, machine learning algorithms can be employed for identifying the target limb of motor intention effectively. As a typical machine learning algorithm for motor imagery decoding, the Riemannian-based kernel support vector machine (RK-SVM) algorithm is not capable of feature extraction from multiple frequency bands, which limits its performance. To solve this problem, the Filter Bank Riemannian-based Kernel Support Vector Machine (FBRK-SVM) method that combines the filter bank structure and Riemannian-based kernel was proposed. In comparative experiments on two commonly used public datasets, it is found that the proposed algorithm can yield higher decoding performance, which provides a new option for the classification of motor imagery.

Yarım Bant Nyquist Filtre ile Sinyallerin Ayrıştırılması ve Gürültünün Giderilmesi Üzerine Bir Araştırma

Medical signal processing is often used for analyzing and detecting diseases from bio-signals. Segmentation, filtering, and noise removal are application examples in this area. The segmentation or decomposition of signals is investigated in this study. Each branch of two band quadrature-mirror orthogonal filter bank structure (analysis and synthesis parts are cascaded) is a half band Nyquist filter when down-up samplers are removed. To perform partitioning, one branch of the filter bank was adapted to a component of input signal to suppress the other element and filter out the target part of the signal. This was done by obtaining the filter-tap weights, which minimize - norm of error at the output. This goal was achieved by employing the exhaustive search method. The filter performance was tested for three scenarios: filtering with and without down-up samplers and wavelet de-noising using one-level decomposition. The comparison was done with Daubechies and Symlet filters. The approach was run and tested for separating synthetically generated evoked potential from noise, and the results show that the designed filter achieves lower total absolute error than the classical wavelet filters.

Filter Bank Sinc-ShallowNet with EMD-based Mixed Noise Adding Data Augmentation for Motor Imagery Classification.

Motor imagery-based brain computer interface (MI-BCI) is a representative active BCI paradigm which is widely employed in the rehabilitation field. In MI-BCI, a classification model is built to identify the target limb from MI-based EEG signals, but the performance of models cannot meet the demand for practical use. Lightweight neural networks in deep learning methods are used to build high performance models in MI-BCI. Small sample sizes and the lack of multi-scale information extraction in frequency domain limit the performance improvement of lightweight neural networks. To solve these problems, the Filter Bank Sinc-ShallowNet (FB-Sinc-ShallowNet) algorithm combined with the mixed noise adding method based on empirical mode decomposition (EMD) was proposed. The FB-Sinc-ShallowNet algorithm improves a lightweight neural network Sinc-ShallowNet with a filter bank structure corresponding to four sensory motor rhythms. The mixed noise adding method employs the EMD method to improve the quality of generated data. The proposed method was evaluated on the BCI competition IV IIa dataset and can achieve highest average accuracy of 77.2%, about 6.34% higher than state-of-the-art method Sinc-ShallowNet. This work implies the effectiveness of filter bank structure in lightweight neural networks and provides a novel option for data augmentation and classification of MI-based EEG signals, which can be applied in the rehabilitation field for decoding MI-EEG with few samples.

Improving the potential of Enhanced Teager Energy Cepstral Coefficients (ETECC) for replay attack detection

In the scope of voice biometrics, the term replay attack, (RA) refers to the dishonest attempt made by an impostor to spoof someone else’s identity by replaying the subject’s previously recorded speech close to the Automatic Speaker Verification (ASV) system under attack. State-of-the-art strategies for RA detection, such as the Enhanced Teager Energy Cepstral Coefficients (ETECC), have shown promising results due to their precision in measuring energy from high frequency components of speech, as a function of two recently defined concepts: signal mass and Enhanced Teager Energy Operator (ETEO). Nevertheless, since the replay mechanism prominently deteriorates the speech signal spectrum just in those spectral zones, we propose the association of ETEO with different strategies to further improve the previous results in getting effective countermeasures against RAs. Specifically, comprehensive evaluations which include a detailed mathematical analysis, a simulation on amplitude and frequency modulated (AM–FM) signals, and a spectrographic inspection involving different filterbank structures, along with their experimental results, are provided in this paper. In addition, ETEO-derived features are contrasted to existing feature sets by using Paraconsistent Feature Engineering (PFE) for feature ranking, expanding our previously published results. Lastly, experiments are performed with ASVSpoof-2017 version 2.0 dataset, Realistic Replay Attack Microphone Array Speech Corpus (ReMASC), BTAS-2016, dataset, ASVSpoof-2019 challenge dataset, and ASVSpoof-2015 challenge dataset, considering Gaussian Mixture Models (GMMs), Convolutional Neural Networks (CNNs) and Light-CNN architectures as being the classifiers. The standalone ETECC-GMM system showed the best performance by producing equal error rates (EERs) of 5.55% and 10.75% on development and evaluation sets, respectively.

Self-tuning variational mode decomposition

Variational mode decomposition (VMD) has attracted a lot of attention recently owing to its robustness to sampling frequency and its high-frequency resolution. However, its performance highly depends on two key preset parameters (the mode number K and the penalty parameter α), both of which tightly limit its adaptability and applications. In this study, a self-tuning VMD (SVMD) is proposed to tackle this problem. Within the proposed method, K and α update themselves respectively and adaptively via the energy ratio and orthogonality between modes in the frequency domain. The proposed SVMD is similar to a matching pursuit method and it shows a VMD-like equivalent filter bank structure but with much less mode-mixing probability. We show that SVMD is more robust to both changes in α and noise level than the original VMD; also, it has better convergence and reduces mode-mixing and end-effect. The experiments on SVMD indicate that SVMD outmatches several classic signal decomposition algorithms. In the end, real-world applications in three fields, namely, length of day variation analysis in geophysics, climate cycle study in meteorology, and oscillation detection in process control, are provided to demonstrate the effectiveness and advantages of the proposed SVMD.

Hardware-efficient auto-reconfigurable hearing aids using 3-level octave interpolated filters for auditory compensation applications.

A reconfigurable hearing aid is a generic type that can be used for various hearing disabilities without modifying the device hardware. This requires several trials to identify the best matching with the impaired person’s audiogram. The objective of this paper is to propose a novel reconfigurable hearing aid of low complexity with auto-adapting capability which makes it suitable for different types of hearing disabilities ranging from mild to severe intensities. The audio spectrum is divided into three regions and for each region, four different schemes are proposed. An automatic selection of the optimum scheme is proposed for all the regions based on hearing thresholds. Octave and fractional interpolation techniques are performed on a Parks-McClellan based prototype filter to generate the various sub-bands in the reconfigurable filter bank structure. The proposed structure uses only 18 coefficient multipliers which save up to 92% of multipliers when compared to other designs. The delay and matching errors are within the globally accepted limits. The hardware implementation executed on Xilinx Kintex-7 FPGA development board has reaffirmed that the structure is compact and power-efficient. The proposed auto-reconfigurable structure can be used for various types of hearing impairments and can avoid the manual interventions for the selection of schemes in audiogram matching. This in turn minimizes the time to establish the best match with the audiogram. Since the proposed structure has minimal complexity, cost-effective implementation of the device is also possible.

Deriving structures of FFT based filter banks by signal flow graphs

The design of filter banks, which is a frequently used operator for many practical applications, has been widely reported in the literature. The requirements of using extensive mathematical manipulations for finding computationally efficient structures of filter banks have imposed severe difficulties for those who lack sufficient mathematical background. This paper is to report procedures of deriving computationally efficient filter bank structures by using a few graphic transformations of equivalent multi-rate operations. It is seen that these procedures used in the derivation process are elegantly simple and straight forwards in concept. In particular, these procedures can be flexibly used to find the computationally efficient structures of both maximally decimated (p=1) and non-maximally decimated (p>1) filter banks, where p is the oversampling factor. For the non-maximally decimated filter banks, the computational complexity needed by our derived structure is less than that needed by the structures reported in the literature.