Denoising Performance Research Articles

An Effective PDE-based Thresholding for MRI Image Denoising and H-FCM-based Segmentation

Image denoising and segmentation play a crucial role in computer graphics and computer vision. A good image-denoising method must effectively remove noise while preserving important boundaries. Various image-denoising techniques have been employed to remove noise, but complete elimination is often impossible. In this paper, we utilize Partial Differential Equation (PDE) and generalised cross-validation (GCV) within Adaptive Haar Wavelet Transform algorithms to effectively denoise an image, with the digital image serving as the input. After denoising, the image is segmented using the Histon-related fuzzy c-means algorithm (H-FCM), with the processed image serving as the output. The proposed method is tested on images exposed to varying levels of noise. The performance of image denoising and segmentation techniques is evaluated using metrics such as Peak Signal-to-Noise Ratio (PSNR) of 77.42, Mean Squared Error (MSE) of 0.0011, and Structural Similarity Index (SSIM) of 0.7848. Additionally, segmentation performance is measured with a sensitivity of 99%, specificity of 98%, and an accuracy of 98%. The results demonstrate that the proposed methods outperform conventional approaches in these metrics. The implementation of the proposed methods is carried out on the MATLAB platform.

Subject-aware PET Denoising with Contrastive Adversarial Domain Generalization.

Recent advances in deep learning (DL) have greatly improved the performance of positron emission tomography (PET) denoising performance. However, DL model performance can vary a lot across subjects, due to the large variability of the count levels and spatial distributions. A generalizable DL model that mitigates the subject-wise variations is highly expected toward a reliable and trustworthy system for clinical application. In this work, we propose a contrastive adversarial learning framework for subject-wise domain generalization (DG). Specifically, we configure a contrastive discriminator in addition to the UNet-based denoising module to check the subject-related information in the bottleneck feature, while the denoising module is adversarially trained to enforce the extraction of subject-invariant features. The sampled low-count realizations from the list-mode data are used as anchor-positive pairs to be close to each other, while the other subjects are used as negative samples to be distributed far away. We evaluated on 97 18F-MK6240 tau PET studies, each having 20 noise realizations with 25% fractions of events. Training, validation, and testing were implemented using 1400, 120, and 420 pairs of 3D image volumes in a subject-independent manner. The proposed contrastive adversarial DG demonstrated superior denoising performance than conventional UNet without subject-wise DG and cross-entropy-based adversarial DG.

Lightweight Infrared Image Denoising Method Based on Adversarial Transfer Learning.

A lightweight infrared image denoising method based on adversarial transfer learning is proposed. The method adopts a generative adversarial network (GAN) framework and optimizes the model through a phased transfer learning strategy. In the initial stage, the generator is pre-trained using a large-scale grayscale visible light image dataset. Subsequently, the generator is fine-tuned on an infrared image dataset using feature transfer techniques. This phased transfer strategy helps address the problem of insufficient sample quantity and variety in infrared images. Through the adversarial process of the GAN, the generator is continuously optimized to enhance its feature extraction capabilities in environments with limited data. Moreover, the generator structure incorporates structural reparameterization technology, edge convolution modules, and progressive multi-scale attention block (PMAB), significantly improving the model's ability to recognize edge and texture features. During the inference stage, structural reparameterization further optimizes the network architecture, significantly reducing model parameters and complexity and thereby improving denoising efficiency. The experimental results of public and real-world datasets demonstrate that this method effectively removes additive white Gaussian noise from infrared images, showing outstanding denoising performance.

STORM image denoising and information extraction.

Stochastic optical reconstruction microscopy (STORM) is extensively utilized in the fields of cell and molecular biology as a super-resolution imaging technique for visualizing cells and molecules. Nonetheless, the imaging process of STORM is frequently susceptible to noise, which can significantly impact the subsequent image analysis. Moreover, there is currently a lack of a comprehensive automated processing approach for analyzing protein aggregation states from a large number of STORM images. This paper initially applies our previously proposed denoising algorithm, UNet-Att, in STORM image denoising. This algorithm was constructed based on attention mechanism and multi-scale features, showcasing a remarkably efficient performance in denoising. Subsequently, we propose a collection of automated image processing algorithms for the ultimate feature extractions and data analyses of the STORM images. The information extraction workflow effectively integrates automated methods of image denoising, objective image segmentation and binarization, and object information extraction, and a novel image information clustering algorithm specifically developed for the morphological analysis of the objects in the STORM images. This automated workflow significantly improves the efficiency of the effective data analysis for large-scale original STORM images.

MbDriver: identifying driver microbes in microbial communities based on time-series microbiome data.

Alterations in human microbial communities are intricately linked to the onset and progression of diseases. Identifying the key microbes driving these community changes is crucial, as they may serve as valuable biomarkers for disease prevention, diagnosis, and treatment. However, there remains a need for further research to develop effective methods for addressing this critical task. This is primarily because defining the driver microbe requires consideration not only of each microbe's individual contributions but also their interactions. This paper introduces a novel framework, called mbDriver, for identifying driver microbes based on microbiome abundance data collected at discrete time points. mbDriver comprises three main components: (i) data preprocessing of time-series abundance data using smoothing splines based on the negative binomial distribution, (ii) parameter estimation for the generalized Lotka-Volterra (gLV) model using regularized least squares, and (iii) quantification of each microbe's contribution to the community's steady state by manipulating the causal graph implied by gLV equations. The performance of nonparametric spline-based denoising and regularized least squares estimation is comprehensively evaluated on simulated datasets, demonstrating superiority over existing methods. Furthermore, the practical applicability and effectiveness of mbDriver are showcased using a dietary fiber intervention dataset and an ulcerative colitis dataset. Notably, driver microbes identified in the dietary fiber intervention dataset exhibit significant effects on the abundances of short-chain fatty acids, while those identified in the ulcerative colitis dataset show a significant correlation with metabolism-related pathways.

Adversarial example denoising and detection based on the consistency between Fourier-transformed layers

Deep neural networks (DNNs) have achieved considerable success in a variety of tasks, and this success has led to an increase in the importance of the security and robustness of algorithms where DNNs are applied. In fact, previous studies have identified that neural networks are vulnerable to adversarial examples that are generated by adding a small amount of noise to the input images. Detection-based methods that can detect adversarial examples among input images are a popular defense technique against adversarial examples. However, these methods reject the detected adversarial images, making it difficult to use the input image for further analysis. In this study, we propose a novel denoising-based detection method that simultaneously denoises and detects an input image when defending against adversarial examples. Our method primarily analyzes images in the feature maps of the target classifier layers based on the consistency between the Fourier domains. We first generate adversarial examples and train the denoising network by minimizing the similarity between the non-adversarial and denoised adversarial images in the feature maps of the Fourier-transformed layers. The predicted denoised images are then input to the adversarial example detector, which determines whether the images are adversarial by exploring the characteristics of the concatenated input and denoised images. There are three main contributions in this paper: (1) We propose a novel denoising-based detection method that simultaneously denoises and detects an input image when defending against adversarial examples. (2) We improve the denoising method for adversarial examples that reconstructs adversarial examples based on the Fourier-transformed feature maps. (3) We improve the detection method by using denoised images in addition to the input images. We tested our proposed method against the fast gradient sign method, basic iterative method, projected gradient descent, Deepfool, and Carlini and Wagner adversarial attack algorithms on two Canadian Institute for Advanced Research (CIFAR) datasets, i.e., CIFAR-10 and CIFAR-100. Our method, which employed a frequency-based denoising mechanism in combination with a detection mechanism, demonstrated significant accuracy improvements in the denoising and detection performances when compared to other state-of-the-art methods.

Pseudo-ISP: Learning pseudo in-camera signal processing pipeline from a color image denoiser

Existing deep denoisers usually ignore the problem of noise discrepancy between training and test images caused by the different noise modeling of sensor and in-camera signal processing (ISP) pipeline, which inevitably degrades the denoising performance. In this paper, we present an unpaired learning scheme to adapt the color image denoisers for handling test images with noise discrepancy. To this end, we consider a practical training setting equipped with a pre-trained color denoiser, a set of test noisy sRGB images, and an unpaired set of clean sRGB images. Then, we propose to alternate between two modules, i.e., Pseudo-ISP training for learning noise modeling of sensor and in-camera signal processing (ISP) pipeline and denoiser adaption. Instead of modeling the complex noise in sRGB images, we assume the signal-dependent and spatially independent noise in rawRGB space, and specifically design the Pseudo-ISP to learn the pseudo ISP pipeline and pseudo rawRGB noise model jointly. Applying the learned Pseudo-ISP to the unpaired set of clean sRGB images, the corresponding realistic noisy sRGB observations can be easily obtained. Conversely, using the generated paired data for model fine-tuning, the color denoiser can be well adapted for test images. Through iterating between Pseudo-ISP training and denoiser adaption, denoising performance of deep denoisers can be further improved. Experiments show that our Pseudo-ISP not only can boost simple Gaussian blurring-based denoiser to achieve competitive performance against CBDNet, but also is effective in improving state-of-the-art deep denoisers, e.g., CBDNet and RIDNet.

Adaptive dual-domain filtering for random seismic noise removal

Random noise significantly reduces the signal-to-noise ratio of seismic data and affects the accuracy of seismic interpretation. Traditional denoising methods typically require manual parameter tuning to increase robustness and accuracy across various random noise levels. In this study, based on the statistical definition of random noise, we use the variance of random noise as the level of random noise and develop an adaptive dual-domain filter (ADDF). The ADDF method estimates the random noise variance in the seismic data and uses this estimation to effectively denoise the seismic data. First, we use a difference operator in two directions to remove the useful structures from the seismic data. The processed data are then used to estimate the global random noise variance through iterative statistical processing. In the denoising process of the ADDF, the seismic data are masked by a bilateral filter in the spatial domain, followed by a short-time Fourier transform with wavelet shrinkage in the frequency domain, both controlled by the adaptively estimated random noise variance. The dual-domain filter is applied iteratively for the best performance. Synthetic experiments demonstrate the robustness of the ADDF in accurately estimating the noise variance without tuning parameters, and its superior denoising performance is evident in synthetic examples and field data when compared with two typical denoising methods: f- x deconvolution and curvelet domain thresholding. As an adaptive random noise estimation and removal method, ADDF relies only on seismic data, making denoising random noise more objective and accurate without manual adjustment.

Unsupervised Domain Adaptation for EM Image Denoising with Invertible Networks.

Electron microscopy (EM) image denoising is critical for visualization and subsequent analysis. Despite the remarkable achievements of deep learning-based non-blind denoising methods, their performance drops significantly when domain shifts exist between the training and testing data. To address this issue, unpaired blind denoising methods have been proposed. However, these methods heavily rely on image-to-image translation and neglect the inherent characteristics of EM images, limiting their overall denoising performance. In this paper, we propose the first unsupervised domain adaptive EM image denoising method, which is grounded in the observation that EM images from similar samples share common content characteristics. Specifically, we first disentangle the content representations and the noise components from noisy images and establish a shared domain-agnostic content space via domain alignment to bridge the synthetic images (source domain) and the real images (target domain). To ensure precise domain alignment, we further incorporate domain regularization by enforcing that: the pseudo-noisy images, reconstructed using both content representations and noise components, accurately capture the characteristics of the noisy images from which the noise components originate, all while maintaining semantic consistency with the noisy images from which the content representations originate. To guarantee lossless representation decomposition and image reconstruction, we introduce disentanglement-reconstruction invertible networks. Finally, the reconstructed pseudo-noisy images, paired with their corresponding clean counterparts, serve as valuable training data for the denoising network. Extensive experiments on synthetic and real EM datasets demonstrate the superiority of our method in terms of image restoration quality and downstream neuron segmentation accuracy. Our code is publicly available at https://github.com/sydeng99/DADn.

Self-supervised multi-echo point cloud denoising in snowfall

Snowfall can cause noise to light detection and ranging (LiDAR) data. This is a problem since it is used in many outdoor applications, e.g., autonomous driving. We propose the task of multi-echo denoising, where the goal is to pick the echo that represents the objects of interest and discard other echoes. Thus, the idea is to pick points from alternative echoes unavailable in standard strongest echo point clouds. Intuitively, we are trying to see through the snowfall. We propose a novel self-supervised deep learning method and the characteristics similarity regularization to achieve this goal. The characteristics similarity regularization utilizes noise characteristics to increase performance. The experiments with a real-world multi-echo snowfall dataset prove the efficacy of multi-echo denoising and superior performance to the baseline. Moreover, based on extensive experiments on a semi-synthetic dataset, our method achieves superior performance compared to the state-of-the-art in self-supervised snowfall denoising. Our work enables more reliable point cloud acquisition in snowfall. The code is available at https://github.com/alvariseppanen/SMEDen.

DeCoGAN: MVCT image denoising via coupled generative adversarial network

Objective. In helical tomotherapy, image-guided radiotherapy employs megavoltage computed tomography (MVCT) for precise targeting. However, the high voltage of megavoltage radiation introduces substantial noise, significantly compromising MVCT image clarity. This study aims to enhance MVCT image quality using a deep learning-based denoising method. Approach. We propose an unpaired MVCT denoising network using a coupled generative adversarial network framework (DeCoGAN). Our approach assumes that a universal latent code within a shared latent space can reconstruct any given pair of images. By employing an encoder, we enforce this shared-latent space constraint, facilitating the conversion of low-quality (noisy) MVCT images into high-quality (denoised) counterparts. The network learns the joint distribution of images from both domains by leveraging samples from their respective marginal distributions, enhanced by adversarial training for effective denoising. Main Results. Compared to an analytical algorithm (BM3D) and three deep learning-based methods (RED-CNN, WGAN-VGG and CycleGAN), the proposed method excels in preserving image details and enhancing human visual perception by removing most noise and retaining structural features. Quantitative analysis demonstrates that our method achieves the highest peak signal-to-noise ratio and Structural Similarity Index Measurement values, indicating superior denoising performance. Significance. The proposed DeCoGAN method shows remarkable MVCT denoising performance, making it a promising tool in the field of radiation therapy.

A Water Environment-Based Simulated Method for Ultrasonic Testing of Slag Inclusion Weld Defects Based on Improved VMD.

The identification of slag inclusion defects in welds is of the utmost importance in guaranteeing the integrity, safety, and prolonged service life of welded structures. Most research focuses on different kinds of weld defects, but branch research on categories of slag inclusion material is limited and critical for safeguarding the quality of engineering and the well-being of personnel. To address this issue, we design a simulated method using ultrasonic testing to identify the inclusion of material categories in austenitic stainless steel. It is based on a simulated experiment in a water environment, and six categories of cubic specimens, including four metallic and two non-metallic materials, are selected to simulate the slag materials of the inclusion defects. Variational mode decomposition optimized by particle swarm optimization is employed for ultrasonic signals denoising. Moreover, the phase spectrum of the denoised signal is utilized to extract the phase characteristic of the echo signal from the water-slag specimen interface. The experimental results show that our method has the characteristics of appropriate decomposition and good denoising performance. Compared with famous signal denoising algorithms, the proposed method extracted the lowest number of intrinsic mode functions from the echo signal with the highest signal-to-noise ratio and lowest normalized cross-correlation among all of the comparative algorithms in signal denoising of weld slag inclusion defects. Finally, the phase spectrum can ascertain whether the slag inclusion is a thicker or thinner medium compared with the weld base material based on the half-wave loss existing or not in the echo signal phase.

Adaptive noise-aware denoising network: Effective denoising for CT images with varying noise intensity

X-ray computed tomography (CT) plays a crucial role in modern medical imaging for its non-invasive acquisition of anatomical information about the human body. However, its inherent carcinogenic risk remains an unavoidable topic. Low-dose CT (LDCT) reduces ionizing radiation exposure to minimize harm to the human body, but it introduces noise and artifacts of varying intensity, which can affect diagnosis and analysis. In recent years, deep learning has demonstrated competitive performance in medical image denoising. However, most current deep learning-based methods primarily focus on specific noise intensities, leading to degradation when faced with noise of different intensities. In this paper, we propose an Adaptive Noise-Aware Denoising Generative Adversarial Networks (ANAD-GAN) to address the aforementioned issues. The proposed noise-aware memory module can remember and decode high-level hidden features of noise of different intensities adaptively through self-updating feature vectors. To explore a more accurate perception of noise distribution, an intricately designed block named BLOGS is incorporated in our framework. We also integrate a discriminator-based detail-preserving loss in our framework to further enhance the visual quality of the denoised images. In summary, our innovative network dynamically adjusts to varying noise intensities, yielding remarkable denoising results. It attains a PSNR of 48.36 and SSIM of 0.9848 across 5936 images in the AAPM dataset, achieves a PSNR of 48.98 and SSIM of 0.9882 on the Philips dataset with 5000 images, and demonstrates outstanding performance with a PSNR of 50.74 and SSIM of 0.9906 on the private Siemens dataset comprising 4000 images. Through extensive experiments on three datasets, our method surpasses various widely adopted techniques in effectively addressing noise of different intensities.

DDT-Net: Dose-Agnostic Dual-Task Transfer Network for Simultaneous Low-Dose CT Denoising and Simulation.

Deep learning (DL) algorithms have achieved unprecedented success in low-dose CT (LDCT) imaging and are expected to be a new generation of CT reconstruction technology. However, most DL-based denoising models often lack the ability to generalize to unseen dose data. Moreover, most simulation tools for LDCT typically operate on proprietary projection data, which is generally not accessible without an established collaboration with CT manufacturers. To alleviate these issues, in this work, we propose a dose-agnostic dual-task transfer network, termed DDT-Net, for simultaneous LDCT denoising and simulation. Concretely, the dual-task learning module is constructed to integrate the LDCT denoising and simulation tasks into a unified optimization framework by learning the joint distribution of LDCT and NDCT data. We approximate the joint distribution of continuous dose level data by training DDT-Net with discrete dose data, which can be generalized to denoising and simulation of unseen dose data. In particular, the mixed-dose training strategy adopted by DDT-Net can promote the denoising performance of lower-dose data. The paired dataset simulated by DDT-Net can be used for data augmentation to further restore the tissue texture of LDCT images. Experimental results on synthetic data and clinical data show that the proposed DDT-Net outperforms competing methods in terms of denoising and generalization performance at unseen dose data, and it also provides a simulation tool that can quickly simulate realistic LDCT images at arbitrary dose levels.

Cotton seed cultivar identification based on the fusion of spectral and textural features.

The mixing of cotton seeds of different cultivars and qualities can lead to differences in growth conditions and make field management difficult. In particular, except for yield loss, it can also lead to inconsistent cotton quality and poor textile product quality, causing huge economic losses to farmers and the cotton processing industry. However, traditional cultivar identification methods for cotton seeds are time-consuming, labor-intensive, and cumbersome, which cannot meet the needs of modern agriculture and modern cotton processing industry. Therefore, there is an urgent need for a fast, accurate, and non-destructive method for identifying cotton seed cultivars. In this study, hyperspectral images (397.32 nm-1003.58 nm) of five cotton cultivars, namely Jinke 20, Jinke 21, Xinluzao 64, Xinluzao 74, and Zhongmiansuo 5, were captured using a Specim IQ camera, and then the average spectral information of seeds of each cultivar was used for spectral analysis, aiming to estab-lish a cotton seed cultivar identification model. Due to the presence of many obvious noises in the < 400 nm and > 1000 nm regions of the collected spectral data, spectra from 400 nm to 1000 nm were selected as the representative spectra of the seed samples. Then, various denoising techniques, including Savitzky-Golay (SG), Standard Normal Variate (SNV), and First Derivative (FD), were applied individually and in combination to improve the quality of the spectra. Additionally, a successive projections algorithm (SPA) was employed for spectral feature selection. Based on the full-band spectra, a Partial Least Squares-Discriminant Analysis (PLS-DA) model was established. Furthermore, spectral features and textural features were fused to create Random Forest (RF), Convolutional Neural Network (CNN), and Extreme Learning Machine (ELM) identification models. The results showed that: (1) The SNV-FD preprocessing method showed the optimal denoising performance. (2) SPA highlighted the near-infrared region (800-1000 nm), red region (620-700 nm), and blue-green region (420-570 nm) for identifying cotton cultivar. (3) The fusion of spectral features and textural features did not consistently improve the accuracy of all modeling strategies, suggesting the need for further research on appropriate modeling strategies. (4) The ELM model had the highest cotton cultivar identification accuracy, with an accuracy of 100% for the training set and 98.89% for the test set. In conclusion, this study successfully developed a highly accurate cotton seed cultivar identification model (ELM model). This study provides a new method for the rapid and non-destructive identification of cotton seed cultivars, which will help ensure the cultivar consistency of seeds used in cotton planting, and improve the overall quality and yield of cotton.

STCC-Filter: A space-time-content correlation-based noise filter with self-adjusting threshold for event camera

Bio-inspired event cameras have become a new paradigm of image sensors detecting illumination changes asynchronously and independently for each pixel. However, their sensitivity to noise degrades the output quality. Most existing denoising methods based on spatiotemporal correlation deteriorate in low light conditions due to frequently bursting noise. To tackle this challenge and remove noise for neuromorphic cameras, this paper proposes space–time-content correlation (STCC) and a novel noise filter with self-adjusted threshold, STCC-Filter. In the proposed denoising algorithm, content correlation is modeled based on the brightness change patterns caused by moving objects. Furthermore, space–time and content support from a sequence of events within the range specified by the threshold which can be programmed based on the real application scenarios are fully utilized to improve the robustness and performance of denoising. STCC-Filter is evaluated on widely used datasets and our labeled synthesized datasets. The experimental results demonstrate that the proposed method outperforms traditional spatiotemporal-correlation-based methods in removing more noise and preserving more signals.

Image Denoise Model Based on Structural Re-Parameterized Uniformer Transformer and UNet

Image denoising is a fundamental problem in computer vision (CV). Vision Transformer (ViT) is an improvement in CV after convolutional neural networks (CNNs). Recently, it has been demonstrated that the RepVGG network of the structural re-parameterization performs well in image tasks. Experiments indicate that the ViT-based uniformer transformer network has successfully balanced local and global information. As image denoising tasks require more local and global information about the image, a novel image denoising model named structural Re-parameterization Uniformer Transformer-UNet (Rep-UUNet) is proposed in this paper. The model is structural and re-parameterized the structure of the Uniformer Transformer network and uses the UNet skip connections to reconstruct the output image. To complete image downstream tasks, the RepVGG model which utilizes the image local information is used. Indicators such as PSNR, SSIM and others are used to assess the performance of image denoising. Experimental results demonstrate that our Rep-UUNet network model outperforms other five models.

New non-local mean methods for MRI denoising based on global self-similarity between values

Magnetic resonance imaging (MRI) is a non-invasive medical imaging technique that provides high-resolution 3D images and valuable insights into human tissue conditions. Even at present, the refinement of denoising methods for MRI remains a crucial concern for improving the quality of the images. This study aims to improve the prefiltered rotationally invariant non-local principal component analysis (PRI-NL-PCA) algorithm. We relaxed the original restrictions using particle swarm optimization to determine optimal parameters for the PCA part of the original algorithm. In addition, we adjusted the prefiltered rotationally invariant non-local mean (PRI-NLM) part by traversing the signal intensities of voxels instead of their spatial positions to reduce duplicate calculations and expand the search volume to the whole image when estimating voxels’ signal intensities. The new method demonstrated superior denoising performance compared to the original approach. Moreover, in most cases, the new algorithm ran faster. Furthermore, our proposed method can also be applied to process Gaussian noise in natural images and has the potential to enhance other NLM-based denoising algorithms.

Generation model meets swin transformer for unsupervised low-dose CT reconstruction

Computed tomography (CT) has evolved into an indispensable tool for clinical diagnosis. Reducing radiation dose crucially minimizes adverse effects but may introduce noise and artifacts in reconstructed images, affecting diagnostic processes for physicians. Scholars have tackled deep learning training instability by exploring diffusion models. Given the scarcity of clinical data, we propose the unsupervised image domain score generation model (UISG) for low-dose CT reconstruction. During training, normal-dose CT images are utilized as network inputs to train a score-based generative model that captures the prior distribution of CT images. In the iterative reconstruction, the initial CT image is obtained using a filtered back-projection algorithm. Subsequently, diffusion-based prior, high-frequency convolutional sparse coding prior, and data-consistency steps are employed to obtain the high-quality reconstructed image. Given the global characteristics of noise, the score network of the diffusion model utilizes a swin transformer structure to enhance the model’s ability to capture long-range dependencies. Furthermore, convolutional sparse coding is applied exclusively to the high-frequency components of the image, to prevent over-smoothing or the loss of crucial anatomical details during the denoising process. Quantitative and qualitative results indicate that UISG outperforms competing methods in terms of denoising and generalization performance.

Cross noise level PET denoising with continuous adversarial domain generalization

Objective. Performing positron emission tomography (PET) denoising within the image space proves effective in reducing the variance in PET images. In recent years, deep learning has demonstrated superior denoising performance, but models trained on a specific noise level typically fail to generalize well on different noise levels, due to inherent distribution shifts between inputs. The distribution shift usually results in bias in the denoised images. Our goal is to tackle such a problem using a domain generalization technique. Approach. We propose to utilize the domain generalization technique with a novel feature space continuous discriminator (CD) for adversarial training, using the fraction of events as a continuous domain label. The core idea is to enforce the extraction of noise-level invariant features. Thus minimizing the distribution divergence of latent feature representation for different continuous noise levels, and making the model general for arbitrary noise levels. We created three sets of 10%, 13%–22% (uniformly randomly selected), or 25% fractions of events from 97 18F-MK6240 tau PET studies of 60 subjects. For each set, we generated 20 noise realizations. Training, validation, and testing were implemented using 1400, 120, and 420 pairs of 3D image volumes from the same or different sets. We used 3D UNet as the baseline and implemented CD to the continuous noise level training data of 13%–22% set. Main results. The proposed CD improves the denoising performance of our model trained in a 13%–22% fraction set for testing in both 10% and 25% fraction sets, measured by bias and standard deviation using full-count images as references. In addition, our CD method can improve the SSIM and PSNR consistently for Alzheimer-related regions and the whole brain. Significance. To our knowledge, this is the first attempt to alleviate the performance degradation in cross-noise level denoising from the perspective of domain generalization. Our study is also a pioneer work of continuous domain generalization to utilize continuously changing source domains.