Denoising Model Research Articles

Data-driven denoising in spinal cord fMRI with principal component analysis.

Numerous approaches have been used to denoise spinal cord functional magnetic resonance imaging (fMRI) data. Principal component analysis (PCA)-based techniques, which derive regressors from a noise region of interest (ROI), have been used in both brain (e.g., CompCor) and spinal cord fMRI. However, spinal cord fMRI denoising methods have yet to be systematically evaluated. Here, we formalize and evaluate a PCA-based technique for deriving nuisance regressors for spinal cord fMRI analysis (SpinalCompCor). In this method, regressors are derived with PCA from a noise ROI, an area defined outside of the spinal cord and cerebrospinal fluid. A parallel analysis is used to systematically determine how many components to retain as regressors for modeling; this designated a median of 11 regressors across three fMRI datasets: motor task (n=26), breathing task (n=27), and resting state (n=10). First-level fMRI modeling demonstrated that principal component regressors did fit noise (e.g., physiological noise from blood vessels), particularly in the resting state fMRI dataset. However, group-level motor task activation maps themselves did not show a clear benefit from including SpinalCompCor regressors over our original denoising model. The potential for collinearity of principal component regressors with the task may be a concern, and this should be considered in future implementations for which task-correlated noise is anticipated.

Leveraging overfitting for good - Two-Step Deep Image Prior Model For Seismic Denoising

Accurate separation of signal and noise constitutes a fundamental prerequisite for achieving high-resolution seismic imaging. A notable recent advancement in this domain is the Deep Image Prior (DIP), an unsupervised deep learning method leveraging deep neural networks (DNNs). The success of this approach lies in the adoption of autoencoders that enables the adaptive extraction of high-fidelity data features. However, establishing an optimal balance between noise suppression and signal preservation remains a non-trivial challenge for DIP-based seismic denoising methods, which is affected by the potential issue of overfitting. This arises from the inappropriate selection of a network architecture and the corresponding hyperparameters, especially the number of training epochs, which strongly influence the learning capacity and feature extraction capabilities of the model. In response to this challenge, we introduce Two-Step DIP (TSDIP), a novel denoising method that exploits overfitting to enhance seismic data quality. In the initial stage, the proposed DNNs are intentionally trained to overfit by effectively attenuating high-frequency noise from the input data. Subsequently, the proposed DNNs are employed iteratively to suppress any residual noise in the newly processed data without damaging useful signal. The overfitting in the first step helps precondition the data to be at a lower noise level while preserving as much as fine-scale features in the signal. To employ an optimal balance, we carefully determine an ideal number of epochs, which is consistently applied in both denoising steps. To assess the effectiveness of the TSDIP method, we present the test results derived from 3D synthetic and field seismic datasets. Our analysis indicates that TSDIP effectively reduces strong noise while preserving key seismic details through the use of overfitting.

Efficient Monte Carlo simulation of streamer discharges with deep-learning denoising models

Efficient Monte Carlo simulation of streamer discharges with deep-learning denoising models

Stimulating Diffusion Model for Image Denoising via Adaptive Embedding and Ensembling.

Image denoising is a fundamental problem in computational photography, where achieving high perception with low distortion is highly demanding. Current methods either struggle with perceptual quality or suffer from significant distortion. Recently, the emerging diffusion model has achieved state-of-the-art performance in various tasks and demonstrates great potential for image denoising. However, stimulating diffusion models for image denoising is not straightforward and requires solving several critical problems. For one thing, the input inconsistency hinders the connection between diffusion models and image denoising. For another, the content inconsistency between the generated image and the desired denoised image introduces distortion. To tackle these problems, we present a novel strategy called the Diffusion Model for Image Denoising (DMID) by understanding and rethinking the diffusion model from a denoising perspective. Our DMID strategy includes an adaptive embedding method that embeds the noisy image into a pre-trained unconditional diffusion model and an adaptive ensembling method that reduces distortion in the denoised image. Our DMID strategy achieves state-of-the-art performance on both distortion-based and perception-based metrics, for both Gaussian and real-world image denoising.

ICESat-2 data denoising and forest canopy height estimation using Machine Learning

Supervised classification methods can distinguish between noise and signal in ice, cloud, and land elevation satellite-2 (ICESat-2) data across various feature perspectives and autonomously optimize parameters. Nevertheless, model generalization remains a significant limitation for practical applications. This study focuses on developing a universal denoising model for ICESat-2 using machine learning algorithms and analyzing its spatial transferability under various forest and terrain conditions. A photon-denoising feature parameter system is developed based on the analysis of the three-dimensional distribution of photons in forested regions. This system reduces the parameters dependent on absolute physical quantities and increases those that are less influenced by terrain and forest features to enhance the model’s transferability. Subsequently, automated machine learning algorithms (AutoML) are used for model selection and parameter optimization across six non-parametric regression models. We evaluate the accuracies of the local, global, and transfer models in estimating canopy height across four representative forested areas in China. Results show that the algorithm can effectively distinguish between signal and noise photons. The estimated canopy heights from signal photons are highly consistent with heights obtained using airborne laser scanning (ALS), exhibiting a Pearson correlation coefficient (r) of 0.89, root mean square errors (RMSE) of 3.75 m, relative root mean square error (rRMSE) of 0.27, relative bias (rBias) of −0.11 and mean Bias of −1.45 m. Notably, the accuracy of canopy height estimation by the global model has increased by an average of 21 % compared to ICESat-2 land-vegetation along-track products (ATL08). Furthermore, the model exhibits significant spatial transfer capabilities, with the accuracies of the transfer model exceeding those of ATL08 by margins ranging from 4 % to 41 %. This study marks a significant advancement in photon-denoising methodologies, providing a robust and transferable solution for large-scale environmental data analysis.

Frequency information enhanced half instance normalization network for denoising electrocardiograms

BackgroundElectrocardiogram (ECG) is crucial in diagnosing and preventing heart diseases. However, its efficacy is compromised by the interference of the external environment, leading to potential misdiagnoses. Thus, it is crucial to remove the noise in ECGs. Recently, deep-learning based ECGs denoising approaches have achieved impressive performance, however, they only considered the time-domain information of ECGs. MethodsIn this work, we propose a Frequency Information Enhanced Half Instance Normalization Network (FIEHINet) which integrates knowledge of both time and frequency domains into a deep-learning model for ECG signal denoising. Two branches are employed to extract time and frequency features for noise eliminating, respectively. Then the ECG signals are reconstructed based on the fused features. Furthermore, masked signal training is introduced to improve the generalization ability. ResultsIn order to evaluate the proposed method, ECGs used are chosen from five different databases. The proposed method for ECG signal denoising achieved Sum of Squared Distances scores of 3.95 ± 7.04, 2.04 ± 3.20, and 0.998 ± 1.579 for three kinds of noise intensities. Meanwhile, the classification experimental results of the processed dataset with the proposed method are 3.8 % higher in F1 score than the original dataset. ConclusionA model for removing mixed noises is successfully developed and tested. SignificanceThis study presents an ECG denoising technique based on Half Instance Normalization, time–frequency information, and masked signal training, which can improve ECG interpretation and potentially reduce misdiagnoses in clinical practice.

UNet-Att: a self-supervised denoising and recovery model for two-photon microscopic image

Two-photon microscopy is indispensable in cell and molecular biology for its high-resolution visualization of cellular and molecular dynamics. However, the inevitable low signal-to-noise conditions significantly degrade image quality, obscuring essential details and complicating morphological analysis. While existing denoising methods such as CNNs, Noise2Noise, and DeepCAD serve broad applications in imaging, they still have limitations in preserving texture structures and fine details in two-photon microscopic images affected by complex noise, particularly in sophisticated structures like neuronal synapses. To improve two-photon microscopy image denoising effectiveness, by experimenting on real two-photon microscopy images, we propose a novel deep learning framework, the UNet-Att model, which integrates a specifically tailored UNet++ architecture with attention mechanisms. Specifically, this approach consists of a sophisticated downsampling module for extracting hierarchical features at varied scales, and an innovative attention module that strategically emphasizes salient features during the integration process. The architecture is completed by an ingenious upsampling pathway that reconstructs the image with high fidelity, ensuring the retention of textural integrity. Additionally, the model supports end-to-end training, optimizing its denoising efficacy. The UNet-Att model proves to surpass mainstream algorithms in the dual objectives of denoising and preserving the textural intricacies of original images, which is evidenced by an increase of 9.42 dB in the high Peak Signal-to-Noise Ratio (PSNR) coupled with an improvement of 0.1131 in the Structural Similarity Index Measurement (SSIM). The ablation experiments reveal the effectiveness of each module. The associated Python packages and datasets of UNet-Att are freely available at https://github.com/ZjjDh/UNet-Att.

Underwater Acoustic Signal LOFAR Spectrogram Denoising Based on Enhanced Simulation

In complex marine environments, extracting target features from acoustic signal is very difficult, making the targets hard to be recognized. Therefore, it is necessary to perform denoising method on the acoustic signal to highlight the target features. However, training deep learning denoising models requires a large mount of acoustic data with labels and obtaining labels with real measured data is also extremely difficult. In this paper, an enhanced simulation algorithm, which considers integrating features of target line spectrum and ocean environmental noise, is proposed to construct a large-scale training sample set. Additionally, a deep convolutional denoising model is presented, which is first train on simulated data and directly applied to real measured data for denoising, enabling line spectrum to be significantly displayed in the time-frequency spectrogram. The results on simulation experiments and sea trials demonstrate that the proposed method can significantly reduce ocean noise while preserving the characteristics of target line spectrum. Furthermore, the experiments demonstrate that the proposed convolutional denoising model has transferability and generalization, making it suitable for denoising underwater acoustic signal in different marine areas.

MoCoDiff: Momentum context diffusion model for low-dose CT denoising

Low-Dose Computed Tomography (LDCT) has gradually replaced Normal-Dose Computed Tomography (NDCT) due to its lower radiation exposure. However, the reduction in radiation dose has led to increased noise and artifacts in LDCT images. To date, many methods for LDCT denoising have emerged, but they often struggle to balance denoising performance with reconstruction efficiency. This paper presents a novel Momentum Context Diffusion model for low-dose CT denoising, termed MoCoDiff. First, MoCoDiff employs a Mean-Preserving Stochastic Degradation (MPSD) operator to gradually degrade NDCT to LDCT, effectively simulating the physical process of CT degradation and greatly reducing sampling steps. Furthermore, the stochastic nature of the MPSD operator enhances the diversity of samples in the training space and calibrates the deviation between network inputs and time-step embedded features. Second, we propose a Momentum Context (MoCo) strategy. This strategy uses the most recent sampling result from each step to update the context information, thereby narrowing the noise level gap between the sampling results and the context data. This approach helps to better guide the next sampling step. Finally, to prevent issues such as over-smoothing of image edges that can arise from using the mean square error loss function, we develop a dual-domain loss function that operates in both the image and wavelet domains. This approach leverages wavelet domain information to encourage the model to preserve structural details in the images more effectively. Extensive experimental results show that our MoCoDiff model outperforms competing methods in both denoising and generalization performance, while also ensuring fast training and inference.

Low Dose CT Image Denoising: A Comparative Study of Deep Learning Models and Training Strategies

Article Low Dose CT Image Denoising: A Comparative Study of Deep Learning Models and Training Strategies Heng Zhao 1, Like Qian 1, Yaqi Zhu 1 and Dingcheng Tian 1,2,∗ 1 Research Institute for Medical and Biological Engineering, Ningbo University, Ningbo 315211, China 2 College of Medicine and Biological Information Engineering, Northeastern University, Shenyang 110016, China ∗ Correspondence: 2310520@stu.neu.edu.cn Received: 8 August 2024; Revised: 10 October 2024; Accepted: 14 October 2024; Published: 5 November 2024 Abstract: Low-dose computed tomography (LDCT) denoising is an important topic in CT image research. Compared with normal-dose CT images, LDCT can reduce the radiation dose of X-rays, decreasing the radiation burden on the human body, which is beneficial to human health. However, quantum noise caused by low-dose rays will reduce the quality of CT images, thereby decreasing the accuracy of clinical diagnosis. In recent years, deep learning-based denoising methods have shown promising advantages in this field. Researchers have proposed some optimized models for low-dose CT image denoising. These methods have enhanced the application of low-dose CT image denoising from different aspects. From the perspective of experimental research, this paper investigates and evaluates some top deep learning models proposed in the field of low-dose image denoising in recent years, with the aim of determining the best models and training strategies for this task. We conducted experiments on seven deep learning models (REDCNN, EDCNN, QAE, OCTNet, UNet, WGAN, CTformer) on the AAPM dataset and the Piglet dataset. Our research shows that UNet has the best denoising effect among the models, obtaining PSNR = 33.06 (AAPM dataset) and PSNR = 31.21 (Piglet dataset), and good generalization capacity is also observed. However, UNet has a large number of parameters, and the time it takes to process an image is about 8 ms, while EDCNN takes about 4.8 ms to process an image, and its average PSNR is ranked second after UNet. EDCNN strikes a balance between denoising performance and processing efficiency, making it ideal for low-dose CT image denoising tasks.

CoaddNet: Enhancing signal-to-noise ratio in single-shot images using convolutional neural networks with coadded image effect

Noise in astronomical images significantly impacts observations and analyses. Traditional denoising methods, such as increasing exposure time and image stacking, are limited when dealing with single-shot images or studying rapidly changing astronomical objects. To address this, we developed a novel deep-learning denoising model, CoaddNet, designed to improve the image quality of single-shot images and enhance the detection of faint sources. To train and validate the model, we constructed a dataset containing high and low signal-to-noise ratio (SNR) images, comprising coadded and single-shot types. CoaddNet combines the efficiency of convolutional operations with the advantages of the Transformer architecture, enhancing spatial feature extraction through a multi-branch structure and reparameterization techniques. Performance evaluation shows that CoaddNet surpasses the baseline model, NAFNet, by increasing the Peak Signal-to-Noise Ratio (PSNR) by 0.03 dB and the Structural Similarity Index (SSIM) by 0.005 while also improving throughput by 35.18%. The model significantly improves the SNR of single-shot images, with an average increase of 22.8, surpassing the noise reduction achieved by stacking 70-90 images. By boosting the SNR, CoaddNet significantly enhances the detection of faint sources, enabling SExtractor to detect an additional 22.88% of faint sources. Meanwhile, CoaddNet reduced the Mean Absolute Percentage Error (MAPE) of flux measurements for detected sources by at least 27.74%.

An efficient noise reduction technique in underwater acoustic signals using enhanced optimization-based residual recurrent neural network with novel loss function

An important process for underwater acoustic signal is noise reduction. In ocean exploration and the military, the minimization of noise in the underwater environment is essential to provide a significant impact in our society. While considering the different acoustic channels and complexity of marine environment, the noise reduction process in acoustic signals is always difficult. Owing to these complexities, an advanced noise reduction mechanism for underwater acoustic signal denoising is performed using adaptive deep learning method is developed. The proposed model involves finding and analyzing the noises present in the underwater acoustic signal, which is helpful for underwater target detection, recognition and acoustic communication quality. Here, a Advanced Recurrent Neural Network with Novel Loss Function (ARRNN-NLF) is implemented for reducing noises in underwater acoustic signals. Hence, the input signal is given to the reduction process where the ARRNN-NLF network is utilized for densification. The high-order nonlinear features from the original signal are extracted and converted into subvectors with fixed lengths based on temporal dimension. Finally, the denoised signal is obtained from the developed ARRNN with a minimum loss between the predicted and target output. Here, the parameters from ARRNN-NLF are optimized by the Enhanced Osprey Optimization Algorithm (EOOA) for enhancing the denoising model. The resultant results are evaluated by diverse conventional denoising models to prove efficiency.

Residual encoder-decoder based architecture for medical image denoising

AbstractHigh-resolution computed tomography (CT) scans require high doses of X-rays, posing potential health risks to patients, including genetic damage and cancer. Conversely, low doses of X-rays result in noise and artifacts in the reconstructed CT scans. Consequently, the problem of denoising low-dose CT (LDCT) images has become a critical yet challenging issue in the field of CT imaging. However, existing deep learning-based LDCT image denoising methods frequently result in the loss of high-frequency features, such as edges and textures, due to the use of mean squared error loss. To address this issue, we propose a method based on high-frequency feature learning to enhance the denoising performance of existing models. Our method is designed to simultaneously learn the primary task of LDCT image denoising and the auxiliary task of LDCT edge detection, thereby improving the denoising performance without increasing the number of model parameters and the inference time. Our method significantly improves the denoising performance of the RED-CNN model, achieving competitive results compared to state-of-the-art denoising models on the AAPM and Qin-LUNG-CT datasets.

Whole-body PET image denoising for reduced acquisition time.

A reduced acquisition time positively impacts the patient's comfort and the PET scanner's throughput. AI methods may allow for reducing PET acquisition time without sacrificing image quality. The study aims to compare various neural networks to find the best models for PET denoising. Our experiments consider 212 studies (56,908 images) for 7MBq/kg injected activity and evaluate the models using 2D (RMSE, SSIM) and 3D (SUVpeak and SUVmax error for the regions of interest) metrics. We tested 2D and 2.5D ResNet, Unet, SwinIR, 3D MedNeXt, and 3D UX-Net. We have also compared supervised methods with the unsupervised CycleGAN approach. The best model for PET denoising is 3D MedNeXt. It improved SSIM on 38.2% and RMSE on 28.1% in 30-s PET denoising and on 16.9% and 11.4% in 60-s PET denoising when compared to the original 90-s PET reducing at the same time SUVmax discrepancy dispersion.

Study of Acoustic Emission Signal Noise Attenuation Based on Unsupervised Skip Neural Network.

Acoustic emission (AE) technology, as a non-destructive testing methodology, is extensively utilized to monitor various materials' structural integrity. However, AE signals captured during experimental processes are often tainted with assorted noise factors that degrade the signal clarity and integrity, complicating precise analytical evaluations of the experimental outcomes. In response to these challenges, this paper introduces an unsupervised deep learning-based denoising model tailored for AE signals. It juxtaposes its efficacy against established methods, such as wavelet packet denoising, Hilbert transform denoising, and complete ensemble empirical mode decomposition with adaptive noise denoising. The results demonstrate that the unsupervised skip autoencoder model exhibits substantial potential in noise reduction, marking a significant advancement in AE signal processing. Subsequently, the paper focuses on applying this advanced denoising technique to AE signals collected during the tensile testing of steel fiber-reinforced concrete (SFRC), the tensile testing of steel, and flexural experiments of reinforced concrete beam, and it meticulously discusses the variations in the waveform and the spectrogram of the original signal and the signal after noise reduction. The results show that the model can also remove the noise of AE signals.

Dual Stage Bionic Spatial Convolutional Restoration Model for Image De-noising and De-mosaicking

The first two important steps in the pipeline for processing picture signals are de-noising and de-mosaicking. The joint solution of the highly uncertain inverse problem of de-noising and de-mosaicking has garnered increased attention in research today. It is difficult to restore high-quality images from raw data in low light because of a variety of disturbances brought on by a low photon count and a complex image signal processing scheme. Even if some restoration and improvement techniques have been used, they might not work in harsh situations, including raw data imaging with brief exposure. Therefore, this research focuses on developing a de-mosaicking and de-noising model with effective end to end manner outcomes. Initially, the pre-processing is conducted using Gaussian filtering to eliminate artifacts from the input image, thereby enhancing the image quality. Then, the proposed method incorporates an Enhanced Spatial Convolutional Residual Net (EnConvResNet) for image de-mosaicking and an Adaptive U-net restoration model for image de-noising. An enhanced gazelle optimization (EnGa) algorithm is used to fine-tune the hyper-parameters of the model in order to maximize its performance and improve its generalization capacity. The proposed method accomplished peak signal to noise ratio (PSNR) and structural similarity index measure (SSIM) of 46.65 and 98.89, respectively.

Acoustic signal adversarial augmentation for pressure pipeline leakage detection

Pressure pipelines are prone to leakage under harsh working condition for a long time, and the leakage detection reaches unsatisfactory performance due to influence of background noise and insufficient sample for acoustic signals. Therefore, the acoustic signals adversarial augmentation method is proposed for pressure pipeline leakage detection based on noise reduction and sample generation. By deeply connecting with generative adversarial network (GAN), denoising autoencoder (DAE) and residual network (ResNet), the adversarial denoising and generation model (ADGM) is established to reduce the noise of acoustic signal. In addition, the trained DAE of ADGM is applied to augment the acoustic samples, thereby completing adversarial augmentation of acoustic signal, which is significant for pressure pipeline leakage detection. Besides, the pipeline leakage experiment is implemented to validate the proposed method on noise reduction and sample generation, which can reach pressure pipeline detection accuracy of 93.02% based on augmented acoustic signal. Further, the effectiveness and superiority of the proposed method are tested by ablation experiments and comparative methods.

MSD: Multi-Order Semantic Denoising Model for Session-Based Recommendations

Session-based recommendations which aim to predict subsequent user–item interactions based on historical user behaviour during anonymous sessions can be challenging to carry out. Two main challenges need to be addressed and improved: (1) how does one analyze these sessions to accurately and completely capture users’ preferences, and (2) how does one identify and eliminate any interference caused by noisy behavior? Existing methods have not adequately addressed these issues since they either neglect the valuable insights that can be gained from analyzing consecutive groups of items or fail to take these noisy data in sessions seriously and handle them properly, which can jointly impede recommendation systems from capturing users’ real intentions. To address these two problems, we designed a multi-order semantic denoising (MSD) model for session-based recommendations. Specifically, we grouped items of different lengths into varying multi-order semantic units to mine the user’s primary intentions from multiple dimensions. Meanwhile, a novel denoising network was designed to alleviate the interference of noisy behavior and provide a more precise session representation. The results of extensive experiments on three real-world datasets demonstrated that the proposed MSD model exhibited improved performance compared with existing state-of-the-art methods in session-based recommendations.

SimNFND: A Forward-Looking Sonar Denoising Model Trained on Simulated Noise-Free and Noisy Data

Given the propagation characteristics of sound waves and the complexity of the underwater environment, denoising forward-looking sonar image data presents a formidable challenge. Existing studies often add noise to sonar images and then explore methods for its removal. This approach neglects the inherent complex noise in sonar images, resulting in inaccurate evaluations of traditional denoising methods and poor learning of noise characteristics by deep learning models. To address the lack of high-quality data for FLS denoising model training, we propose a simulation algorithm for forward-looking sonar data based on RGBD data. By utilizing rendering techniques and noise simulation algorithms, high-quality noise-free and noisy sonar data can be rapidly generated from existing RGBD data. Based on these data, we optimize the loss function and training process of the FLS denoising model, achieving significant improvements in noise removal and feature preservation compared to other methods. Finally, this paper performs both qualitative and quantitative analyses of the algorithm’s performance using real and simulated sonar data. Compared to the latest FLS denoising models based on traditional methods and deep learning techniques, our method demonstrates significant advantages in denoising capability. All inference results for the Marine Debris Dataset (MDD) have been made open source, facilitating subsequent research and comparison.

PostFocus: automated selective post-acquisition high-throughput focus restoration using diffusion model for label-free time-lapse microscopy.

High-throughput time-lapse imaging is a fundamental tool for efficient living cell profiling at single-cell resolution. Label-free phase-contrast video microscopy enables noninvasive, nontoxic, and long-term imaging. The tradeoff between speed and throughput, however, implies that despite the state-of-the-art autofocusing algorithms, out-of-focus cells are unavoidable due to the migratory nature of immune cells (velocities >10 μm/min). Here, we propose PostFocus to (i) identify out-of-focus images within time-lapse sequences with a classifier, and (ii) deploy a de-noising diffusion probabilistic model to yield reliable in-focus images. De-noising diffusion probabilistic model outperformed deep discriminative models with a superior performance on the whole image and around cell boundaries. In addition, PostFocus improves the accuracy of image analysis (cell and contact detection) and the yield of usable videos. Open-source code and sample data are available at: https://github.com/kwu14victor/PostFocus.