Noise Robustness Research Articles

Robustness Comparison of Traditional and Deep Learning-Based Image Segmentation Models on Noisy Images

Abstract— Detecting objects in images with precision becomes increasingly difficult when the input is degraded by noise or distortion, as is common in real-world applications. This study investigates and compares the performance of a traditional edge detection technique (Canny) and a deep learning-based model (U- Net) for binary image segmentation under noisy conditions. We simulate three types of image noise—Gaussian, Salt & Pepper, and Motion Blur—on a public dataset and evaluate both models using Dice Score, Intersection over Union (IoU), and Structural Similarity Index (SSIM) to assess their quantitative accuracy and perceptual consistency. While U-Net demonstrates stronger resilience in capturing complex structures and maintaining segmentation accuracy, Canny proves more computationally efficient and surprisingly stable under certain distortions. Our results highlight the importance of selecting segmentation methods based not only on accuracy but also on noise robustness and deployment constraints. This work offers practical insights into the trade-offs between traditional and deep learning models for vision tasks in noisy environments. Keywords— Image Segmentation, U-Net, Canny Edge Detection, Noise Robustness, Dice Score, IoU, SSIM, Deep Learning, Computer Vision

Enhanced RSSD-based Unknown Radio Emitter Localization Via Weighted Grey Correlation Degree

Abstract Accurate tracking of indoor unknown radio emitter (URE) is a key task in the field of wireless communication security. Fingerprint positioning methods based on received signal strength difference (RSSD) perform well on URE localization challenges with unknown transmitting power and frequency. Since the noise level of RSSD parameters is amplified compared to received signal strength, the matching accuracy between RSSD signal features and geographical locations is reduced. This paper presents a RSSD-based positioning algorithm using improved singular value decomposition (SVD) and weighted grey correlation degree (GCD), called ISVD-WGCD, to increase noise robustness and positioning accuracy. Firstly, local least squares curve fitting is applied to optimize the selection of singular value truncation point to reduce the noise level of RSSD offline database. Secondly, combining the offline RSSD singular value variation coefficient and online RSSD standard deviation, an AP weight allocation scheme is designed to set the trust degrees of different APs reasonably. Finally, a RSSD-based weighted GCD method is proposed by AP weight allocation to improve the matching accuracy between offline and online RSSDs, and to select the optimal reference points. Compared with the conventional weighted K-nearest neighbors (WKNN), SVD-KNN, Bayes, convolutional neural network (CNN) and FUZZY-GREY algorithms, simulation and experimental results show that the proposed method can obtain superior positioning performance under different AP numbers, grid distances and noise levels.

Gearbox Fault Diagnosis Under Noise and Variable Operating Conditions Using Multiscale Depthwise Separable Convolution and Bidirectional Gated Recurrent Unit with a Squeeze-and-Excitation Attention Mechanism

Gearbox condition monitoring is essential for ensuring the reliability of power transmission systems. However, the existing methods are constrained by shallow feature extraction and unidirectional temporal modeling. To address these limitations, this study proposes a novel fault diagnosis framework that integrates multiscale depthwise separable convolution, bidirectional gated recurrent units, and a squeeze-and-excitation attention mechanism. This approach enables multiscale feature extraction from vibration signals, bidirectional temporal modeling, and the enhancement of critical fault-related information. The experimental results demonstrate that the proposed method significantly outperforms conventional models in terms of fault diagnosis accuracy, noise robustness, and adaptability to varying operating conditions. The attention mechanism effectively suppresses noise interference, while bidirectional temporal modeling accurately captures fault propagation characteristics, thereby improving adaptability to dynamic conditions. This research provides a highly robust solution for intelligent gearbox fault diagnosis in complex industrial environments.

Rolling bearings fault diagnosis using a one-dimensional vision transformer with multi-scale residual convolution

Abstract Traditional transformer-based fault diagnosis methods exhibit inherent limitations in extracting local features, making it challenging to effectively identify critical fault characteristics when processing rolling bearing vibration signals, especially in high-noise environments. Therefore, this paper proposes an intelligent diagnosis method based on noise learning, called the multi-scale residual convolution one-dimensional vision transformer (MRC-1DViT). Firstly, the MRC module is constructed using an adaptive noise learning approach to extract local fault features across multiple scales, capturing fault information of varying dimensions, and reduce noise interference. Subsequently, the extracted features are fed into a 1DViT based fault diagnosis model to identify the long-distance dependencies between features. Additionally, contra norm is introduced to mitigate the issue of dimensional collapse in the model. Finally, the effectiveness of MRC-1DViT is validated using the Case Western Reserve University (CWRU) rolling bearing dataset and the Southeast University (SEU) gearbox dataset. Experimental results demonstrate that MRC-1DViT exhibits superior noise robustness and higher stability compared to existing methods. Notably, under a high-noise environment of −6 dB signal-to-noise ratio, the proposed method achieves fault diagnosis accuracies of 91.38% and 89.38%, respectively.

Lightweight concrete crack recognition model based on improved MobileNetV3

This study created the C//Sim attention mechanism employing the parallel connection of the CA attention mechanism and the SimAm attention mechanism to detect cracks in lightweight concrete. MobileNetV3 was improved using the above method, and a lightweight concrete crack recognition model, MobileNetV3-C//Sim, was established. To validate the model’s practicality, this paper has been tested on self-built and public datasets. The improved model performs higher accuracy, recall, precision, and F1 values than Mobilenetv3 in both datasets, with increases of 0.44–0.69% and 0.46–0.89% for the binary and multi classification tasks, respectively. For the CA attention mechanism, SimAm attention mechanism, and ablation tests with different combinations of each other showed that the parallel connection combination was superior to the single-type, front-to-back concatenation combination. In noise testing with different attention mechanisms, the C//Sim reduction is the smallest. It is verified to have better noise immunity and robustness. Regarding the number of model parameters, the proposed method involves only 2.90 M, which is 30.17% less than that of MobileNetV3. The method can provide a model reference for further concrete crack lightweight identification research.

Noise-Robust image edge detection based on multi-scale automatic anisotropic morphological Gaussian Kernels.

This paper presents a novel multi-scale, noise-robust edge detection method that employs multi-scale automatic anisotropic morphological Gaussian kernels to extract edge maps from input images. It addresses the issue of cross-edge detection failure in the Canny edge detector. Compared to other edge detection methods, the proposed approach offers significant advantages in maintaining noise robustness while achieving high edge resolution and accuracy. The paper is structured into five key sections. First, we propose a multi-scale automatic anisotropic morphological directional derivative (AMDD) to capture local gray-level variations around each pixel at multiple scales. Second, a new fused edge strength map (ESM) is introduced based on the multi-scale AMDD. Third, we analyze why the Canny isotropic Gaussian kernel detector fails to detect cross edges. Additionally, the edge contour is extracted by incorporating the fused ESMs and the edge direction map (EDM), which are processed through spatial and directional matching filters, into the standard Canny detection framework. Finally, we evaluate the proposed method using precision-recall (PR) curves and Pratt's Figure of Merit (FOM). We compare its performance with existing state-of-the-art detectors on a standard dataset. Experimental results demonstrate that the proposed method effectively reduces noise, mitigates irrelevant signal interference, and smooths the image, showing competitive performance in edge detection tasks.

Intelligent analysis algorithm for power engineering data based on improved BiLSTM

Traditional BiLSTM (Bidirectional Long Short-Term Memory) has limited processing capabilities for long-time series, insufficient adaptability to the characteristics of power engineering data, and inadequate accuracy in predicting problems in specific application scenarios. This paper introduces an attention mechanism to improve BiLSTM’s ability to capture long time series, integrate the time series of power engineering data, and design an intelligent analysis algorithm that adapts to multi-dimensional features. Experimental results show that in load forecasting, the mean square errors of the improved BiLSTM in summer and winter are 0.02 and 0.025 respectively, and R2 are 0.985 and 0.982 respectively. In equipment fault diagnosis, the accuracy of improved BiLSTM under current, voltage, temperature and pressure is significantly higher than that of models such as GRU.(Gated Recurrent Unit) This paper improves BiLSTM, combines multi-dimensional feature fusion and multi-head self-attention mechanism, and optimizes it according to the characteristics of power engineering data, enhance noise robustness, and reduce noise impact by 8%. Although the training time and memory are slightly increased, the convergence speed is faster. The improved BiLSTM power engineering data analysis algorithm significantly improves the accuracy and robustness of tasks such as power load forecasting and equipment fault diagnosis by introducing self attention mechanism and multi-dimensional feature fusion. It is more adaptable in complex temporal patterns and multi-source data modeling.

Robust Single-Cell RNA-Seq Analysis Using Hyperdimensional Computing: Enhanced Clustering and Classification Methods

Background. Single-cell RNA sequencing (scRNA-seq) has transformed genomics by enabling the study of cellular heterogeneity. However, its high dimensionality, noise, and sparsity pose significant challenges for data analysis. Methods. We investigate the use of Hyperdimensional Computing (HDC), a brain-inspired computational framework recognized for its noise robustness and hardware efficiency, to tackle the challenges in scRNA-seq data analysis. We apply HDC to both supervised classification and unsupervised clustering tasks. Results. Our experiments demonstrate that HDC consistently outperforms established methods such as XGBoost, Seurat reference mapping, and scANVI in terms of noise tolerance and scalability. HDC achieves superior accuracy in classification tasks and maintains robust clustering performance across varying noise levels. Conclusions. These results highlight HDC as a promising framework for accurate and efficient single-cell data analysis. Its potential extends to other high-dimensional biological datasets including proteomics, epigenomics, and transcriptomics, with implications for advancing bioinformatics and personalized medicine.

Quantification of tissue stiffness with magnetic resonance elastography and finite difference time domain (FDTD) simulation-based spatiotemporal neural network.

Quantification of tissue stiffness with magnetic resonance elastography and finite difference time domain (FDTD) simulation-based spatiotemporal neural network.

Towards parameter-free attentional spiking neural networks.

Towards parameter-free attentional spiking neural networks.

Weak fault feature extraction of rolling bearing based on IHBA-SVMD-MCKD and Dynamic Threshold via Pearson and Kurtosis

Abstract Extracting weak fault features under strong background noise interference is key to early fault diagnosis of rolling bearings. A method combining the improved honey badger algorithm (IHBA), successive variational mode decomposition (SVMD), and maximum correlated kurtosis deconvolution (MCKD) is proposed. The SVMD method effectively suppresses noise while retaining key fault features. To improve feature extraction, IHBA is used as the optimization tool, and a composite index PE composed of the Pearson correlation coefficient ratio-envelope spectrum peak factor is used as the fitness function, enhancing the efficiency and stability of parameter optimization. To address the issue of incorrect IMF component selection leading to excessive noise and poor robustness, the Dynamic Threshold via Pearson and Kurtosis is constructed, which adaptively selects IMF components rich in fault features for signal reconstruction. MCKD can highlight continuous pulse signals obscured by noise. IHBA is used again, with a composite index EE composed of envelope entropy-envelope spectrum peak factor as the fitness function, and the SVMD-processed signal is input to MCKD. Finally, envelope demodulation is used to extract rolling bearing fault feature frequencies, significantly improving the accuracy of weak fault feature extraction. Through verification with simulated signals of early rolling bearing faults and the full-lifecycle dataset XJTU-SY, the proposed method improves the accuracy and robustness of early fault diagnosis under strong background noise and other interference factors, providing a possible solution for extracting weak fault features.

Advances in Zeroing Neural Networks: Bio-Inspired Structures, Performance Enhancements, and Applications

Zeroing neural networks (ZNN), as a specialized class of bio-Iinspired neural networks, emulate the adaptive mechanisms of biological systems, allowing for continuous adjustments in response to external variations. Compared to traditional numerical methods and common neural networks (such as gradient-based and recurrent neural networks), this adaptive capability enables the ZNN to rapidly and accurately solve time-varying problems. By leveraging dynamic zeroing error functions, the ZNN exhibits distinct advantages in addressing complex time-varying challenges, including matrix inversion, nonlinear equation solving, and quadratic optimization. This paper provides a comprehensive review of the evolution of ZNN model formulations, with a particular focus on single-integral and double-integral structures. Additionally, we systematically examine existing nonlinear activation functions, which play a crucial role in determining the convergence speed and noise robustness of ZNN models. Finally, we explore the diverse applications of ZNN models across various domains, including robot path planning, motion control, multi-agent coordination, and chaotic system regulation.

Compressed sensing for telehealth assessment of disordered voices

Abstract The goal of this research work is to propose compressed sensing (CS) as a component of a telemedicine system for compressing speech signals in the framework of telehealth assessment of disordered voices. The main challenge is the reconstruction of the compressed medical audio data without affecting the acoustics markers used by the clinicians in the diagnosis and interpretation. Firstly, the speech signal is compressed by applying the CS at the appropriate compression rate, to each frame to reduce the size of the signal. Next, the compressed speech signal is transmitted through a transmission channel. The speech signal is recovered from the CS-based compressed signal using an appropriate algorithm. Finally, the acoustics markers jitter, shimmer, harmonic-to-noise rate (HNR) and cepstral peak prominence (CPP) are computed for disordered voices assessment. The effect of the compression on the acoustic markers is investigated by using a corpus of Spanish sustained vowels /a/ uttered by 79 normophonic and dysphonic speakers with different degrees of dysphonia. The robustness of the proposed method agains background noise is evaluated by corrupting the voice signals with synthetic white Gaussian noise at different levels. Experimental results show that there are no significant differences in the correlations between the auditory evaluation scores and acoustics markers computed from the original speech signal and the recovered signal compressed via CS-based method even at a compression rate (CR) as high as 75%. The performance of the CS-based method in terms of the computed acoustics markers and mean square error (MSE), for a fixed CR, is compared to that of the MP3 and advanced audio coding (AAC) audio compression formats. The CS technique is a very effective tool for compressing speech signals in the framework of telehealth assessment of disordered voices.

Comparison Of Feature Extraction Techniques For Long Short-Term Memory Models In Indonesian Automatic Speech Recognition

Automatic Speech Recognition (ASR) faced challenges in accuracy and noise robustness, particularly in Bahasa Indonesia. This research addressed the limitations of single feature extraction methods, such as Mel-Frequency Cepstral Coefficients (MFCC), which were sensitive to noise, and Relative Spectral Transform - Perceptual Linear Predictive (RASTA-PLP), which was less effective in frequency representation, by proposing a hybrid approach that combined both techniques using Long Short-Term Memory (LSTM) models. MFCC enhanced spectral accuracy, while RASTA-PLP improved noise robustness, resulting in a more adaptive and informative acoustic representation. The evaluation demonstrated that the hybrid method outperformed single and non-extraction approaches, achieving a Character Error Rate (CER) of 0.5245 on clean data and 0.8811 on noisy data, as well as a Word Error Rate (WER) of 0.9229 on clean data and 1.0015 on noisy data. Although the hybrid approach required longer training times and higher memory usage, it remained stable and effective in reducing transcription errors. These findings suggested that the hybrid method was an optimal solution for Indonesian speech recognition in various acoustic conditions.

Image dehazing algorithm based on light-value weighted allocation and multi-layer restricted perception

In image dehazing, the dehazing performance in bright regions and the model’s robustness to noise are critical evaluation criteria. However, existing dehazing models often suffer from distortions in the bright areas and exhibit weak noise resistance. We propose an image dehazing algorithm based on light-value weighted allocation and multi-layer restricted perception (DWARP) to address these issues. The proposed algorithm first constructs an atmospheric light estimation module based on weighted allocation. It reduces the initial estimation error by zeroing out overexposed pixel values, then identifying key factors affecting atmospheric light estimation and assigning weights accordingly. Finally, a threshold-restricted adjustment is applied to the estimated result, achieving a three-stage refinement in atmospheric light estimation accuracy. Secondly, by computing the universal range of transmittance for both bright and non-bright regions, a multi-layer restricted perception scheme for transmittance is designed. This approach transforms the distortion issue in bright regions into a problem of reducing transmittance estimation errors. Finally, to enhance the visual quality of the dehazed image and improve the model’s noise robustness, a brightness adjustment module and a Gaussian denoising module are embedded into the dehazing model. Experimental results demonstrate that the DWARP algorithm effectively prevents distortion in the dehazing process for bright regions and enhances the model’s noise robustness. The DWARP algorithm achieved an average PSNR of 37.41 dB, an average SSIM of 88.74%, and an average VIF of 0.89 across four datasets, with an average dehazing time of 0.633 s. Compared to the RIDCP algorithm, DWARP improved the PSNR by an average of 6.63 dB, SSIM by 2.18%, and VIF by 0.03 while enhancing dehazing efficiency by an average of 0.055 s. To validate the effectiveness of the DWARP algorithm, we conducted dehazing experiments on four foggy datasets, demonstrating the proposed algorithm’s effectiveness and superiority. The DWARP algorithm not only effectively processes bright regions, such as the sky, in foggy images but also exhibits strong noise resistance, validating the scientific and theoretical correctness of the proposed improvements. This dehazing model provides a novel approach for fog removal in fields such as intelligent transportation, contributing to the advancement and development of these areas.

LayerAct: Advanced Activation Mechanism for Robust Inference of CNNs

In this work, we propose a novel activation mechanism called LayerAct for CNNs. This approach is motivated by our theoretical and experimental analyses, which demonstrate that Layer Normalization (LN) can mitigate a limitation of existing activation functions regarding noise robustness. However, LN is known to be disadvantageous in CNNs due to its tendency to make activation outputs homogeneous. The proposed method is designed to be more robust than existing activation functions by reducing the upper bound of influence caused by input shifts without inheriting LN's limitation. We provide analyses and experiments showing that LayerAct functions exhibit superior robustness compared to ElementAct functions. Experimental results on three clean and noisy benchmark datasets for image classification tasks indicate that LayerAct functions outperform other activation functions in handling noisy datasets while achieving superior performance on clean datasets in most cases.

SAP: Corrective Machine Unlearning with Scaled Activation Projection for Label Noise Robustness

Label corruption, where training samples are mislabeled due to non-expert annotation or adversarial attacks, significantly degrades model performance. Acquiring large, perfectly labeled datasets is costly, and retraining models from scratch is computationally expensive. To address this, we introduce Scaled Activation Projection (SAP), a novel SVD (Singular Value Decomposition)-based corrective machine unlearning algorithm. SAP mitigates label noise by identifying a small subset of trusted samples using cross-entropy loss and projecting model weights onto a clean activation space estimated using SVD on these trusted samples. This process suppresses the noise introduced in activations due to the mislabeled samples. In our experiments, we demonstrate SAP’s effectiveness on synthetic noise with different settings and real-world label noise. SAP applied to the CIFAR dataset with 25% synthetic corruption show upto 6% generalization improvements. Additionally, SAP can improve the generalization over noise robust training approaches on CIFAR dataset by ∼ 3.2% on average. Further, we observe generalization improvements of 2.31% for a Vision Transformer model trained on naturally corrupted Clothing1M.

MEDSAGE: Enhancing Robustness of Medical Dialogue Summarization to ASR Errors with LLM-generated Synthetic Dialogues

Automatic Speech Recognition (ASR) systems are pivotal in transcribing speech into text, yet the errors they introduce can significantly degrade the performance of downstream tasks like summarization. This issue is particularly pronounced in clinical dialogue summarization, a low-resource domain where supervised data for fine-tuning is scarce, necessitating the use of ASR models as black-box solutions. Employing conventional data augmentation for enhancing the noise robustness of summarization models is not feasible either due to the unavailability of sufficient medical dialogue audio recordings and corresponding ASR transcripts. To address this challenge, we propose MEDSAGE, an approach for generating synthetic samples for data augmentation using Large Language Models (LLMs). Specifically, we leverage the in-context learning capabilities of LLMs and instruct them to generate ASR-like errors based on a few available medical dialogue examples with audio recordings. Experimental results show that LLMs can effectively model ASR noise, and incorporating this noisy data into the training process significantly improves the robustness and accuracy of medical dialogue summarization systems. This approach addresses the challenges of noisy ASR outputs in critical applications, offering a robust solution to enhance the reliability of clinical dialogue summarization.

Bearing fault diagnosis based on efficient cross space multiscale CNN transformer parallelism

Fault diagnosis of wind turbine bearings is crucial for ensuring operational safety and reliability. However, traditional serial-structured deep learning models often fail to simultaneously extract spatio- temporal features from fault signals in noisy environments, leading to critical information loss. To address this limitation, this paper proposes a Wind Turbine Bearing Fault Diagnosis Model Based on Efficient Cross Space Multiscale CNN Transformer Parallelism (ECMCTP). The model first transforms one-dimensional vibration signals into two-dimensional time-frequency images using Continuous Wavelet Transform (CWT). Subsequently, parallel branches are employed to extract spatio-temporal features: the Convolutional Neural Network (CNN) branch integrates a multiscale feature extraction module, a Reversed Residual Structure (RRS), and an Efficient Multiscale Attention (EMA) mechanism to enhance local and global feature extraction capabilities; the Transformer branch combines Bidirectional Gated Recurrent Units (BiGRU) and Transformer to capture both local temporal dynamics and long-term dependencies. Finally, the features from both branches are concatenated along the channel dimension and classified using a softmax classifier. Experimental results on two publicly available bearing datasets demonstrate that the proposed model achieves 100% accuracy under noise-free conditions and maintains superior noise robustness under low signal-to-noise ratio (SNR) conditions, showcasing excellent robustness and generalization capabilities.

Noise-robust spatial-temporal dynamic graph domain adaptation network for fault diagnosis under varying working conditions

Abstract Modern industrial production relies heavily on rotating machinery, and its failure can lead to significant financial losses and safety hazards. Therefore, the study of intelligent fault diagnosis is of critical importance. Multi-sensor data diagnosis enables the extraction of fault features from multiple perspectives, making it an essential approach to improving the reliability of fault diagnosis models. However, effectively extracting time-domain information and capturing the interrelationships between multiple sensors remain significant challenges. Additionally, in practical industrial production processes, the working conditions of rotating machinery, such as load and rotating speeds, are constantly fluctuating. Thus, we propose a novel noise-robust model, named Spatial-Temporal Dynamic Graph Domain Adaptation Network (STDGDAN), for fault diagnosis under varying working conditions. Firstly, we design a Spatial-Temporal Dynamic Graph Module (STDGM) based on Dynamic Graph Convolutional Networks (DGCNs) and Bidirectional Gated Recurrent Units (BiGRUs). The STDGM leverages BiGRUs to extract time-domain features and employs DGCNs to capture spatial features through gating units. The module combines multi-sensor and spatial-temporal information for diagnosis, resulting in improved noise robustness. Secondly, we introduce a multi-domain, multi-kernel maximum mean discrepancy (MK-MMD) as a domain adaptation strategy. Specifically, the distances between the source and target domains are computed in both the spatial-temporal and frequency domain feature spaces. The sum of these distances forms the domain adaptation loss function. Finally, the effectiveness and superiority of the proposed method are validated through extensive experiments.