Low-frequency Information Research Articles

Metamaterial-Based Analog Recurrent Neural Network Toward Machine Intelligence

Wave-based analog computation platforms hold the promise of realizing intelligent information processing with high speed and low computational cost, but current design principles face intractable problems in processing low-frequency wave information end to end at the physical layer. Here we theoretically demonstrate a concept of a metamaterial-based analog recurrent neural network (Meta-RNN) for intelligent classification of physical information carried by mechanical vibrations. With the spatiotemporal recurrence relation induced by coupled local resonators of the simulated metamaterial model, the Meta-RNN is capable of memorizing the internal hidden state in the dynamic displacement field. The trained coupling and synergy of local resonators lead to desirable vibration energy localization with selective frequency extraction capability, which can distinguish the intrinsic characteristics of vibration information. The proposed Meta-RNN provides a foundation for implementing a promising mechanical analog processor for intelligent vibration information processing and classification, and paves the way to efficient machine learning platforms for machine intelligence.

WDIG: a wavelet domain image generation framework based on frequency domain optimization

In the end-to-end image generation task, the spatial domain of pixel space cannot explicitly separate the low-frequency general information such as texture and color from the high-frequency detail information such as structure and identity. The loss function calculated in the spatial domain fails to effectively constrain the maintenance of detail information, and the generated image quality is insufficient. In this paper, a wavelet domain image generation (WDIG) framework is proposed to preserve the frequency information of images, in which the loss functions are constructed in the pixel space and wavelet space. In the pixel space, the low-frequency and high-frequency characteristic information of the signal are obtained by setting the appropriate Gaussian kernel and adopting the Gaussian fuzzy method. The loss function of ell_{1} norm spatial domain is constructed for the low-frequency and high-frequency characteristic information. In the wavelet space, the corresponding channel sub-band coefficients are obtained by wavelet transform, and the image is explicitly separated into high-frequency information and low-frequency information. The ell_{1} norm frequency domain loss function is constructed respectively for the sub-band coefficients. The WDIG can constrain model training more accurately and optimize model more precisely, so as to better maintain the details and quality of generated image. The WDIG framework is evaluated in the image generation applications including style transfer, image translation and Generative Adversarial Nets (GAN) Inversion. Experimental results show that the WDIG framework can effectively retain the details of images and generate more realistic images, and improve the image quality of the above applications in image generation.

Monetary policy and information shocks in a block-recursive SVAR

This study introduces a new estimator that combines block-recursive restrictions with higher-order moment conditions and non-Gaussian shocks. The proposed estimator improves the accuracy of the estimation, simplifies labeling, and allows for relaxing the independence and non-Gaussianity assumptions in comparison to a purely data-driven approach. We use the approach to disentangle the interaction of stock prices and interest rates into monetary policy and stock market information shocks. We find that traditional monetary policy shocks move interest rates and stock prices in opposite directions, whereas information shocks move both variables in the same direction. Moreover, we utilize high-frequency data from FOMC announcements to derive a proxy for central bank information shocks and show that these shocks are statistically relevant for the low-frequency stock market information shock.

A multi-channel framework based Local Binary Pattern with two novel local feature descriptors for texture classification

Local binary pattern (LBP) has been widely used in various application fields, including object detection, texture analysis, and remote sensing. To improve discrimination performance, many LBP-based methods have been proposed for extracting different local feature information for texture classification. However, current LBP-based algorithms typically describe local features at a single sampling scale, but some significant and discriminative texture feature information is contained between different scales. Therefore, the lack of capability to extract cross-scale features can lead to losing critical texture features. Moreover, low-frequency texture information should be accorded great importance in texture classification. Additionally, if the diversity and validity of local feature extraction are lacking, the effectiveness of classification capability suffers. To address these issues, (1) we propose a completed cross-scale Local binary pattern (ccsLBP) operator to extract cross-scale texture features. (2) A mean-filtered Local binary pattern (LBPmf) operator is presented to highlight the important low-frequency texture information of texture images. (3) We build a high-performing multi-channel framework based Local binary pattern (MC-LBP) for texture classification, which combines complementary features extracted by LBP, ccsLBP, and LBPmf hybridly to form a final feature vector of the texture image. The effectiveness of the proposed MC-LBP framework is verified on 6 representative texture databases, and the experimental results demonstrate its state-of-art texture classification performance.

A borehole porosity prediction method with focusing on local shape

Porosity is a valuable parameter reflecting petroleum storage performance and plays an important role in reservoir exploration. Various studies have confirmed the feasibility of establishing a mapping between geophysical logging and porosity by using machine learning methods. The existing works are generally in the form of depth point-to-point. However, due to the temporal and spatial characteristics of the stratum deposition, the correlation under continuous depths of logging curves should be considered in addition to the absolute values. Attempts have been made to achieve the memory for low-frequency information with continuous depths by introducing the concept of time in recurrent neural networks. This logging porosity prediction strategy has disadvantages due to two reasons: (i) periodicity is present in the logging data, but is weak, and (ii) small-scale information of the logging curves have a guiding meaning on the detail of porosity. In this paper, we present the first attempt to describe the high-frequency variation in the local shape of the logging curves as spatial small-scale features and propose a ConvLSTM-based framework in the porosity prediction task. Specifically, the local shape features of the curves are further extracted by convolutional operations as the contextual information is memorized. We form 1-dimensional (1D) images in a segment-to-point manner (fixing segment to 16 depth points) to make the logging data suitable for convolution operations while increasing the volume of data. Further, segment-to-point samples are automatically intercepted and feed into model with segment-to-segment prediction patterns according to the view window size. Various experiments are carried out on a real dataset from the Shengli Oilfield, Jiyang Depression, and the results validate the effectiveness of ConvLSTM for local shape information extraction and the superiority in porosity prediction.

DVDR-SRGAN: Differential Value Dense Residual Super-Resolution Generative Adversarial Network

In the field of single-image super-resolution reconstruction, GAN can obtain the image texture more in line with the human eye. However, during the reconstruction process, it is easy to generate artifacts, false textures, and large deviations in details between the reconstructed image and the Ground Truth. In order to further improve the visual quality, we study the feature correlation between adjacent layers and propose a differential value dense residual network to solve this problem. We first use the deconvolution layer to enlarge the features, then extract the features through the convolution layer, and finally make a difference between the features before being magnified and the features after being extracted so that the difference can better reflect the areas that need attention. In the process of extracting the differential value, using the dense residual connection method for each layer can make the magnified features more complete, so the differential value obtained is more accurate. Next, the joint loss function is introduced to fuse high-frequency information and low-frequency information, which improves the visual effect of the reconstructed image to a certain extent. The experimental results on Set5, Set14, BSD100, and Urban datasets show that our proposed DVDR-SRGAN model is improved in terms of PSNR, SSIM, and LPIPS compared with the Bicubic, SRGAN, ESRGAN, Beby-GAN, and SPSR models.

Graph Convolution Point Cloud Super-Resolution Network Based on Mixed Attention Mechanism

In recent years, point cloud super-resolution technology has emerged as a solution to generate a denser set of points from sparse and low-quality point clouds. Traditional point cloud super-resolution methods are often optimized based on constraints such as local surface smoothing; thus, these methods are difficult to be used for complex structures. To address this problem, we proposed a novel graph convolutional point cloud super-resolution network based on a mixed attention mechanism (GCN-MA). This network consisted of two main parts, i.e., feature extraction and point upsampling. For feature extraction, we designed an improved dense connection module that integrated an attention mechanism and graph convolution, enabling the network to make good use of both global and local features of the point cloud for the super-resolution task. For point upsampling, we adopted channel attention to suppress low-frequency information that had little impact on the up-sampling results. The experimental results demonstrated that the proposed method significantly improved the point cloud super-resolution performance of the network compared to other corresponding methods.

Spectral Correlation and Spatial High–Low Frequency Information of Hyperspectral Image Super-Resolution Network

Hyperspectral images (HSIs) generally contain tens or even hundreds of spectral segments within a specific frequency range. Due to the limitations and cost of imaging sensors, HSIs often trade spatial resolution for finer band resolution. To compensate for the loss of spatial resolution and maintain a balance between space and spectrum, existing algorithms were used to obtain excellent results. However, these algorithms could not fully mine the coupling relationship between the spectral domain and spatial domain of HSIs. In this study, we presented a spectral correlation and spatial high–low frequency information of a hyperspectral image super-resolution network (SCSFINet) based on the spectrum-guided attention for analyzing the information already obtained from HSIs. The core of our algorithms was the spectral and spatial feature extraction module (SSFM), consisting of two key elements: (a) spectrum-guided attention fusion (SGAF) using SGSA/SGCA and CFJSF to extract spectral–spatial and spectral–channel joint feature attention, and (b) high- and low-frequency separated multi-level feature fusion (FSMFF) for fusing the multi-level information. In the final stage of upsampling, we proposed the channel grouping and fusion (CGF) module, which can group feature channels and extract and merge features within and between groups to further refine the features and provide finer feature details for sub-pixel convolution. The test on the three general hyperspectral datasets, compared to the existing hyperspectral super-resolution algorithms, suggested the advantage of our method.

MSLEFC: A low-frequency focused underwater acoustic signal classification and analysis system

Many feature extraction techniques are employed to study the classification of underwater acoustic, however, most of these techniques result in the loss of detailed information in this field. A Multi-scale short-time Fourier transform (MS-STFT) method was proposed to improve the low-frequency information and maintain the detailed information by increasing the number of channels. An effective data augmentation approach was proposed and a Ladder-like Encode (LE) architecture was built to increase the generalizability of the model and the classification accuracy for the features extracted. Finally, A Frequency-CAM (FC) method is proposed to analyze the frequency band locations that the neural network is interested in when conducting classification tasks on various classes of data. The integration of the above technologies is called MSLEFC. The performance of this system has been tested on two ship radiation noise datasets and obtained 82.94% and 96.06% accuracy, respectively. On the ShipsEar dataset, the proposed method increases the accuracy by 0.7% over the previous state-of-the-art method. The proposed model architecture has significant improvements in parameter size and computation compared to ResNet50.

'Real-life' benefit of hearing preservation cochlear implantation in the paediatric population: a single-site case–control study

IntroductionCochlear implantation with hearing preservation (HPCI) has allowed a cochlear implant (CI) electrode to be implanted while trying to preserve residual acoustic low-frequency hearing. The concept arises from the importance...

Denoising Diffusion Probabilistic Feature-Based Network for Cloud Removal in Sentinel-2 Imagery

Cloud contamination is a common issue that severely reduces the quality of optical satellite images in remote sensing fields. With the rapid development of deep learning technology, cloud contamination is expected to be addressed. In this paper, we propose Denoising Diffusion Probabilistic Model-Cloud Removal (DDPM-CR), a novel cloud removal network that can effectively remove both thin and thick clouds in optical image scenes. Our network leverages the denoising diffusion probabilistic model (DDPM) architecture to integrate both clouded optical and auxiliary SAR images as input to extract DDPM features, providing significant information for missing information retrieval. Additionally, we propose a cloud removal head adopting the DDPM features with an attention mechanism at multiple scales to remove clouds. To achieve better network performance, we propose a cloud-oriented loss that considers both high- and low-frequency image information as well as cloud regions in the training procedure. Our ablation and comparative experiments demonstrate that the DDPM-CR network outperforms other methods under various cloud conditions, achieving better visual effects and accuracy metrics (MAE = 0.0229, RMSE = 0.0268, PSNR = 31.7712, and SSIM = 0.9033). These results suggest that the DDPM-CR network is a promising solution for retrieving missing information in either thin or thick cloud-covered regions, especially when using auxiliary information such as SAR data.

HDGN: Heat diffusion graph network for few-shot learning

A heat diffusion graph network (HDGN) is proposed in this paper, which retains more similar graph signals in the spectral domain, for few-shot learning. Convolution on the graph is essentially the filtering of the graph signal. Most existing graph-network-based few-shot learning methods process graph signals with high-pass filters to get the difference in information. However, the low-frequency similar information is usually more valuable in the few-shot tasks. A joint low-pass filter is constructed to filter low-frequency graph signals, that is, heat kernel convolution aggregates similar information from neighboring nodes. The obtained low-frequency similarity information is utilized to update the representations of nodes on the graph. In addition, a more robust mixed metric is designed to dynamically update the edge feature of the graph. Predicting Unknown Nodes on Graphs by Alternating Updates of Node Representation and Edge Matrix. The experimental results also demonstrate that HDGN achieves better performance for the few-shot classification task.

Information Purification Network for Remote Sensing Image Super-Resolution

Recently, several well-performing deep convolutional neural networks were proposed for remote sensing image super-resolution (SR). However, these methods rarely consider that remote sensing images are corruptible by additional noise, blurring, and other factors. Therefore, to eliminate the interference of these factors, especially the noise, we propose a novel information purification network (IPN) for remote sensing image SR. The proposed information purification block (IPB) can process channel-wise features differently by channel separation and rescale spatial-wise features adaptively through the proposed multi-scale spatial attention mechanism. We further design an information group to explore a more powerful expressive combination of IPBs. Moreover, long and short skip connections can transmit abundant low-frequency information, making IPBs pay more attention to high-frequency information. We mix the images under various degradation models as training data in the training phase. In this way, the network can directly reconstruct various degraded images. Experiments on AID and UC Merced Land-Use datasets under multiple degradation models demonstrate that the proposed IPN performs better than state-of-the-art methods.

(MLE$^{2}$A$^{2}$U)-Net: Image Super-Resolution via Multi-Level Edge Embedding and Aggregated Attentive Upsampler Network

Given a degraded low-resolution input image, super-resolution (SR) aims at restoring the lost textures and structures and generating high-resolution image content. Significant advances in image super-resolution have been made lately, dominated by convolutional neural networks (CNNs). The top performing CNN-based SR networks typically employ very deep models for embracing the benefits of generating spatially precise results, but at the cost of loss of long-term contextual information. Additionally, state-of-the-art (SOTA) methods generally lack in maintaining the balance between spatial details and contextual information, which is the basic requirement for exhibiting superior performance in SR task. For restoration application like SR, the overall network generally demands efficient preservation of low-frequency information and reconstruction of high-frequency details. Thus, our work presents a novel architecture with the holistic objective of maintaining spatially-precise representation by collecting contextual content and restoring multi-frequency information throughout the network. Our proposed model learns an enriched set of features, that besides combining contextual information from multiple scales simultaneously preserves the high-resolution spatial details. The core of our approach is a novel non-local and local attention (NLLA) block which focuses on (1) learning enriched features by collecting information from multiple scales, (2) simultaneously handling the different frequency information, and (3) effectively fusing the relevant low-frequency and high-frequency information by ignoring the redundant features. Additionally, for effectively mapping the low-resolution features to high-resolution, we propose a novel aggregated attentive up-sampler (AAU) block that attentively learns the weights to up-sample the refined low-resolution feature maps to high-resolution output. Extensive experiments on the benchmark SR datasets demonstrate that the proposed method achieves appealing performance, both qualitatively and quantitatively.

Robust deep learning object recognition models rely on low frequency information in natural images.

Machine learning models have difficulty generalizing to data outside of the distribution they were trained on. In particular, vision models are usually vulnerable to adversarial attacks or common corruptions, to which the human visual system is robust. Recent studies have found that regularizing machine learning models to favor brain-like representations can improve model robustness, but it is unclear why. We hypothesize that the increased model robustness is partly due to the low spatial frequency preference inherited from the neural representation. We tested this simple hypothesis with several frequency-oriented analyses, including the design and use of hybrid images to probe model frequency sensitivity directly. We also examined many other publicly available robust models that were trained on adversarial images or with data augmentation, and found that all these robust models showed a greater preference to low spatial frequency information. We show that preprocessing by blurring can serve as a defense mechanism against both adversarial attacks and common corruptions, further confirming our hypothesis and demonstrating the utility of low spatial frequency information in robust object recognition.

Bi-filter multiscale-diversity-entropy-based weak feature extraction for a rotor-bearing system

Multiscale-based entropy methods have proven to be a promising tool for extracting fault information due to their high feature extraction ability and easy application. Despite multiscale analysis showing great potential in extracting fault characteristics, it has some drawbacks, such as cutting the data length and neglecting high-frequency information. This paper proposes a bi-filter multiscale diversity entropy (BMDE) to filter comprehensive fault information and address the data length problem. First, the low-frequency information is filtered out by moving average in a multi-low procedure and the high-frequency information is filtered out by an adjacent subtraction in a multi-high procedure. Second, a modified coarse-grained process is introduced to overcome the issue of data length. The validity of the BMDE method is evaluated using both simulation signals and experimental measurements. Results demonstrate that the proposed method offers optimal feature extraction capability with the highest diagnostic accuracy compared with four other traditional entropy-based diagnosis methods.

Dynamics of Functional Networks for Syllable and Word-Level Processing.

Speech comprehension requires the ability to temporally segment the acoustic input for higher-level linguistic analysis. Oscillation-based approaches suggest that low-frequency auditory cortex oscillations track syllable-sized acoustic information and therefore emphasize the relevance of syllabic-level acoustic processing for speech segmentation. How syllabic processing interacts with higher levels of speech processing, beyond segmentation, including the anatomical and neurophysiological characteristics of the networks involved, is debated. In two MEG experiments, we investigate lexical and sublexical word-level processing and the interactions with (acoustic) syllable processing using a frequency-tagging paradigm. Participants listened to disyllabic words presented at a rate of 4 syllables/s. Lexical content (native language), sublexical syllable-to-syllable transitions (foreign language), or mere syllabic information (pseudo-words) were presented. Two conjectures were evaluated: (i) syllable-to-syllable transitions contribute to word-level processing; and (ii) processing of words activates brain areas that interact with acoustic syllable processing. We show that syllable-to-syllable transition information compared to mere syllable information, activated a bilateral superior, middle temporal and inferior frontal network. Lexical content resulted, additionally, in increased neural activity. Evidence for an interaction of word- and acoustic syllable-level processing was inconclusive. Decreases in syllable tracking (cerebroacoustic coherence) in auditory cortex and increases in cross-frequency coupling between right superior and middle temporal and frontal areas were found when lexical content was present compared to all other conditions; however, not when conditions were compared separately. The data provide experimental insight into how subtle and sensitive syllable-to-syllable transition information for word-level processing is.

Infrared and visible image fusion with entropy-based adaptive fusion module and mask-guided convolutional neural network

Since infrared and visible images have different descriptions of the same scene and have a lot of complementary information, they are usually fused to obtain a more comprehensive image. To overcome the detail information losing, fusion artifacts, and unnatural fusion in the traditional image fusion, we propose a novel algorithm for infrared and visible image fusion. Firstly, we introduce a guided filter to decompose the input image into a first-level base layer, a second-level base layer and a detail layer. Specifically, for the first-level base layer fusion, we propose an entropy-based image adaptive fusion module, which is guided by the image quality and the information entropy. To reduce the low-frequency information losing, we use the maximum absolute value rule to fuse the second-level base layer. In addition, for the detail layer containing many textures and edges, we construct a mask-guided deep convolutional neural network to fuse it. We label the infrared image pixel-by-pixel in the training set to obtain the foreground target mask, and construct the loss function jointly considering the mask, gradient, and the structural similarity of the image. Masks are only used when training the network. Finally, we reconstruct the image. Experiments show that our algorithm has achieved good performance in both subjective and objective evaluation.

IFormerFusion: Cross-Domain Frequency Information Learning for Infrared and Visible Image Fusion Based on the Inception Transformer

The current deep learning-based image fusion methods can not sufficiently learn the features of images in a wide frequency range. Therefore, we proposed IFormerFusion, which is based on the Inception Transformer and cross-domain frequency fusion. To learn features from high- and low-frequency information, we designed the IFormer mixer, which splits the input features through the channel dimension and feeds them into parallel paths for high- and low-frequency mixers to achieve linear computational complexity. The high-frequency mixer adopts a convolution and a max-pooling path, while the low-frequency mixer adopts a criss-cross attention path. Considering that the high-frequency information relates to the texture detail, we designed a cross-domain frequency fusion strategy, which trades high-frequency information between the source images. This structure can sufficiently integrate complementary features and strengthen the capability of texture retaining. Experiments on the TNO, OSU, and Road Scene datasets demonstrate that IFormerFusion outperforms other methods in object and subject evaluations.

Waveform inversion via reduced order modeling

We introduce a novel approach to waveform inversion based on a data-driven reduced order model (ROM) of the wave operator. The presentation is for the acoustic wave equation, but the approach can be extended to elastic or electromagnetic waves. The data are time resolved measurements of the pressure wave gathered by an acquisition system that probes the unknown medium with pulses and measures the generated waves. We propose to solve the inverse problem of velocity estimation by minimizing the square misfit between the ROM computed from the recorded data and the ROM computed from the modeled data, at the current guess of the velocity. We give a step by step computation of the ROM, which depends nonlinearly on the data and yet can be obtained from them in a noniterative fashion, using efficient methods from linear algebra. We also explain how to make the ROM robust to data inaccuracy. The ROM computation requires the full array response matrix gathered with colocated sources and receivers. However, we find that the computation can deal with an approximation of this matrix, obtained from towed-streamer data using interpolation and reciprocity on-the-fly. Although the full-waveform inversion approach of nonlinear least-squares data fitting is challenging without low-frequency information, due to multiple minima of the data fit objective function, we find that the ROM misfit objective function has better behavior, even for a poor initial guess. We also find by explicit computation of the objective functions in a simple setting that the ROM misfit objective function has convexity properties, whereas the least-squares data fit objective function displays multiple local minima.