speckle noise reduction

This paper investigates the statistical limits of nonparametric function estimation under composite noise—specifically, multiplicative speckle noise combined with additive Gaussian noise. Addressing a fundamental gap in existing theory, we establish, for the first time, a minimax lower bound framework in the presence of speckle noise and derive the optimal convergence rate over Hölder classes: ((max{1,sigma_n^4}/n)^{2eta/(2eta+1)}). This result reveals that speckle noise introduces a (sigma_n^4) dependence, causing a substantial slowdown in convergence relative to purely additive noise settings—quantitatively confirming its inherently greater statistical difficulty. Our analysis systematically characterizes the degradation mechanism of statistical efficiency across varying speckle-to-additive noise intensity ratios. The theoretical benchmarks derived herein provide foundational guidance for nonparametric modeling in speckle-dominated imaging modalities, including synthetic aperture radar (SAR) and ultrasound imaging.

Ultrasound speckle noise exhibits tissue dependence and strong spatial correlation, rendering conventional self-supervised denoising methods—such as Noise2Noise and blind-spot networks—ineffective. To address this, we propose the first self-supervised ultrasound denoising framework requiring only a single noisy input image. Our method models multi-scale anatomical consistency and tissue-specific noise variations via multi-scale perturbations, and jointly separates tissue structures from noise by integrating low-rank–sparse decomposition priors with deep neural networks. Crucially, it operates without clean ground-truth labels, paired training data, or repeated acquisitions. Evaluated on both simulated and real carotid ultrasound datasets, our approach significantly outperforms state-of-the-art filtering and learning-based methods in quantitative metrics and visual quality. Moreover, it demonstrates strong generalization across different ultrasound devices, highlighting its clinical practicality and robustness.

This study addresses the challenge in coherent imaging where multiplicative speckle noise renders nonparametric regression functions unidentifiable and undermines conventional least-squares-based deep learning approaches. To overcome this, the authors formulate a joint model incorporating both multiplicative speckle and additive Gaussian noise and propose a likelihood-based deep neural network framework for estimating smooth nonparametric regression functions. They establish, for the first time, a minimax theory for speckle regression, demonstrating that its statistical complexity is comparable to that of purely additive Gaussian noise settings—thereby resolving the identifiability barrier posed by multiplicative noise. In both low-dimensional and sparse high-dimensional regimes, the proposed method achieves minimax-optimal convergence rates up to logarithmic factors, matching those under additive Gaussian noise alone. Both theoretical analysis and numerical experiments confirm the method’s effectiveness in speckle removal and estimation consistency.

To address low detection accuracy and poor noise robustness in underwater sonar image object detection, this paper systematically transfers and adapts nine state-of-the-art optical-image deep denoising models to the sonar domain for the first time, proposing a novel “complementary multi-model denoising + multi-frame collaborative feature fusion” paradigm to overcome limitations of single-model denoising. The method integrates advanced denoisers—including DnCNN and RIDNet—with YOLOv5 and RT-DETR detection frameworks, achieving an average mAP improvement of 12.7% across five public sonar datasets; incorporating multi-frame fusion further boosts performance by 3.2%, significantly outperforming raw input. Key contributions are: (1) the first systematic cross-modal validation of deep denoising model transfer from optical to sonar imaging; and (2) a new architecture jointly optimizing multi-source denoising and multi-frame temporal modeling.

This work addresses the challenge of speckle noise in ultrasound imaging, which degrades image quality and obscures anatomical boundaries. Existing denoising methods often suffer from over-smoothing and limited generalizability. To overcome these limitations, the authors propose a Noise-aware Boundary-preserving Generative Learning (NBGL) framework that jointly models noise level estimation and boundary preservation for the first time. NBGL employs a dual-branch architecture—comprising speckle suppression and boundary enhancement pathways—and introduces a noise-aware weighted Feature-wise Linear Modulation (wFiLM) mechanism to enable adaptive feature fusion. Evaluated on 141 3D transvaginal ultrasound volumes across six noise levels, the method consistently outperforms state-of-the-art approaches, achieving superior denoising performance while significantly enhancing anatomical fidelity.

This work proposes a learning-based statistical refinement method for improving the denoising results of a given denoiser without knowing the precise noise distribution or accessing clean images or calibration data. While there are many existing successful denoising approaches for handling different kinds of noise, they typically require accurate modelling of the images and the noise (implicitly or explicitly), and hence the denoising results can be suboptimal due to different practical factors such as imperfect models, unreliable noise assumptions, or low quality data. In particular, when clean image samples are not available and there is a lack of knowledge of the underlying noise distribution, which is the case in various practical situations, the results may not well align with the noise statistics. The unawareness of the useful statistical information leads to suboptimal results. This work aims to make the best use of the statistical information to improve the consistency between the given denoising results and the noise statistics, under the assumption that the noise is conditionally pixel-wise independent given the clean signal. A method, based on a Bayesian formulation of an auxiliary signal in the noisy data, is proposed for evaluating the consistency of the denoising results, without precise information on noise distribution. By leveraging the statistical information from noisy data, the method enhances the statistical noise consistency and improves denoising quality.

Standard data augmentation techniques exhibit limited efficacy in laser speckle-based material classification because they overlook the structural statistical properties inherent to speckle patterns, which arise from coherent interference. This work proposes a parameterized augmentation framework to systematically evaluate the impact of various perturbations—including rotation, Gaussian blur, independent noise, spatially correlated speckle-aware noise, intensity jitter, and spatial masking—on classification performance. Leveraging ResNet18 and EfficientNet-B0 models alongside ordinary least squares analysis, the study demonstrates that augmentation effectiveness hinges on preserving the spatial and frequency-domain structure of speckle rather than the magnitude of perturbation. Structure-preserving augmentations, such as spatially correlated noise, substantially enhance robustness, whereas Gaussian blur and independent noise degrade performance. The proposed framework accounts for up to 87.9% of performance variance, establishing a design principle centered on physically informed, structure-preserving augmentation.

This work addresses the challenge of denoising ultrasound images, which are inherently corrupted by electronic and speckle noise. Conventional methods rely on fixed noise models, while learning-based approaches require extensive annotated data and suffer from domain shift. To overcome these limitations, the authors propose A2A—a test-time training framework that operates without pretraining or labeled data, leveraging only a single noisy synthetic aperture ultrasound sample. By performing self-supervised contrastive learning in a pyramid latent space, A2A disentangles anatomical structures from noise in a single pass. This is the first method to employ self-contrastive learning within a pyramid latent representation for structure–noise separation, thereby inherently mitigating domain shift. Experiments demonstrate substantial improvements: in simulations, SNR increases by 69.3% and CNR by 34.4%; in vivo cardiac, liver, and kidney imaging achieves 84.8% SNR and 25.7% CNR gains using just two aperture acquisitions.