Diffusion Model Research Articles

Comparative analysis of the distribution of methane emissions in the designated regions based on satellite observations and calculations using transport and diffusion model

Comparative analysis of the distribution of methane emissions in the designated regions based on satellite observations and calculations using transport and diffusion model

A diffusion model multi-scale feature fusion network for imbalanced medical image classification research

Background and objectiveMedicine image classification are important methods of traditional medical image analysis, but the trainable data in medical image classification is highly imbalanced and the accuracy of medical image classification models is low. In view of the above two common problems in medical image classification. This study aims to: (i) effectively solve the problem of poor training effect caused by the imbalance of class imbalanced data sets. (ii) propose a network framework suitable for improving medical image classification results, which needs to be superior to existing methods. MethodsIn this paper, we put in the diffusion model multi-scale feature fusion network (DMSFF), which mainly uses the diffusion generation model to overcome imbalanced classes (DMOIC) on highly imbalanced medical image datasets. At the same time, it is processed according to the cropped image augmentation strategy through cropping (IASTC). Based on this, we use the new dataset to design a multi-scale feature fusion network (MSFF) that can fully utilize multiple hierarchical features. The DMSFF network can effectively solve the problems of small and imbalanced samples and low accuracy in medical image classification. ResultsWe evaluated the performance of the DMSFF network on highly imbalanced medical image classification datasets APTOS2019 and ISIC2018. Compared with other classification models, our proposed DMSFF network achieved significant improvements in classification accuracy and F1 score on two datasets, reaching 0.872, 0.731, and 0.906, 0.836, respectively. ConclusionsOur newly proposed DMSFF architecture outperforms existing methods on two datasets, and verifies the effectiveness of generative model inverse balance for imbalance class datasets and feature enhancement by multi-scale feature fusion. Further, the method can be applied to other class imbalanced data sets where the results will be improved.

Two-dimensional problem for an infinite body of two coaxial cylinders under the action of a solenoidal body force in the theory of thermoelastic diffusion

This study investigates the dynamic response of an infinite structure comprising two coaxial cylinders employing the one-relaxation-time generalized thermoelastic diffusion model. The inner cylinder, serving as a cavity, possessing a radius a, while the outer cylinder has radius b. The inner cylinder is insulated and experiencing a localized thermal load with thickness 2 h. The outer cylinder, with radius b, is also experiencing to a localized thermal load with width 2 h, and its surface exhibits no traction. Additionally, the structure experiences the influence of a solenoidal body force. The analysis of this problem employs a combination of Laplace transform, inverse Laplace transform, and exponential Fourier transform techniques. Numerical analysis of the resulting solutions reveals the variations in temperature, displacement, radial stress, concentration, and potential fields as a function of the radial coordinate. Results indicate a minimal impact of the time interval on the temperature distribution, while displacement and radial stress exhibit significant time-dependent behavior. These findings highlight the significance of incorporating thermoelastic diffusion theory in the analysis of real-world engineering problems.

From text to mask: Localizing entities using the attention of text-to-image diffusion models

Diffusion models have revolted the field of text-to-image generation recently. The unique way of fusing text and image information contributes to their remarkable capability of generating highly text-related images. From another perspective, these generative models imply clues about the precise correlation between words and pixels. This work proposes a simple but effective method to utilize the attention mechanism in the denoising network of text-to-image diffusion models. Without additional training time nor inference-time optimization, the semantic grounding of phrases can be attained directly. We evaluate our method on Pascal VOC 2012 and Microsoft COCO 2014 under weakly-supervised semantic segmentation setting and our method achieves superior performance to prior methods. In addition, the acquired word-pixel correlation is generalizable for the learned text embedding of customized generation methods, requiring only a few modifications. To validate our discovery, we introduce a new practical task called “personalized referring image segmentation” with a new dataset. Experiments in various situations demonstrate the advantages of our method compared to strong baselines on this task. In summary, our work reveals a novel way to extract the rich multi-modal knowledge hidden in diffusion models for segmentation.

MTLSC-Diff: Multitask learning with diffusion models for hyperspectral image super-resolution and classification

Hyperspectral image (HSIs) super-resolution (SR) can improve the spatial resolution of images for subsequent application tasks. In recent years, SR methods based on deep learning have gained widespread attention. However, most of the existing SR methods do not take into account the needs of specific application tasks when designing the network structure. These methods may not be able to efficiently generate high-quality images that satisfy the specific application tasks, leading to degradation of the performance of subsequent application tasks. To solve this problem, we propose a multi-task learning architecture based on the diffusion model, namely MTLSC-Diff. MTLSC-Diff combines the SR network and the classification network in a multi-task learning manner on the basis of the diffusion model. MTLSC-Diff achieves mutual guidance of the two tasks by iterating the image super-resolution and classification tasks, thus gradually reconstructing high-quality images and improving classification accuracy. The guided operations for each time step are performed by the specially designed Mutual-Guidance SR-Classification Synergy Module (M-GSCS). M-GSCS refines the multi-scale image obtained at the previous time step and uses the predicted high spatial resolution image for classification. Meanwhile, a class-guided SR dynamic refinement strategy (C-GSR) is proposed in M-GSCS, which uses multi-scale classification results to guide target scale images to learn new knowledge to further reconstruct high-quality images. Experimental results on relevant datasets show that our method significantly improves the super-resolution performance as well as the classification performance.

ScDiffusion: conditional generation of high-quality single-cell data using diffusion model.

Single-cell RNA sequencing (scRNA-seq) data are important for studying the laws of life at single-cell level. However, it is still challenging to obtain enough high-quality scRNA-seq data. To mitigate the limited availability of data, generative models have been proposed to computationally generate synthetic scRNA-seq data. Nevertheless, the data generated with current models are not very realistic yet, especially when we need to generate data with controlled conditions. In the meantime, diffusion models have shown their power in generating data with high fidelity, providing a new opportunity for scRNA-seq generation. In this study, we developed scDiffusion, a generative model combining the diffusion model and foundation model to generate high-quality scRNA-seq data with controlled conditions. We designed multiple classifiers to guide the diffusion process simultaneously, enabling scDiffusion to generate data under multiple condition combinations. We also proposed a new control strategy called Gradient Interpolation. This strategy allows the model to generate continuous trajectories of cell development from a given cell state. Experiments showed that scDiffusion could generate single-cell gene expression data closely resembling real scRNA-seq data. Also, scDiffusion can conditionally produce data on specific cell types including rare cell types. Furthermore, we could use the multiple-condition generation of scDiffusion to generate cell type that was out of the training data. Leveraging the Gradient Interpolation strategy, we generated a continuous developmental trajectory of mouse embryonic cells. These experiments demonstrate that scDiffusion is a powerful tool for augmenting the real scRNA-seq data and can provide insights into cell fate research. scDiffusion is openly available at the GitHub repository https://github.com/EperLuo/scDiffusion or Zenodo https://zenodo.org/doi/10.5281/zenodo.13268742.

AirECG: Contactless Electrocardiogram for Cardiac Disease Monitoring via mmWave Sensing and Cross-domain Diffusion Model

The electrocardiogram (ECG) has always served as a crucial biomedical examination for cardiac diseases monitoring and diagnosing. Typical ECG measurement requires attaching electrodes to the body, which is inconvenient for long-term monitoring. Recent wireless sensing maps wireless signals reflected from human chest into electrical activities of heart so as to reconstruct ECG contactlessly. While making great progress, we find existing works are effective only for healthy populations with normal ECG, but fall short when confronted with the most desired usage scenario: reconstructing ECG accurately for people with cardiac diseases such as atrial fibrillation, premature ventricular beat. To bridge the gap, we propose AirECG, which moves forward to reconstruct ECG for both healthy people and even cardiac patients with morbid ECG, i.e., irregular rhythm and anomalous ECG waveform, via contactless millimeter-wave sensing. To realize AirECG, we first custom-design a cross-domain diffusion model that can perform multiple iteration denoising inference, in contrast with the single-step generative models widely used in previous works. In this way, AirECG is able to identify and eliminate the distortion due to the unstable and irregular cardiac activities, so as to synthesize ECG even during abnormal beats. Furthermore, we enhance the determinacy of AirECG, i.e., to generate high-fidelity ECG, by designing a calibration guidance mechanism to combat the inherent randomness issue of the probabilistic diffusion model. Empirical evaluation demonstrates AirECG's ability of ECG synthesis with Pearson correlation coefficient (PCC) of 0.955 for normal beats. Especially for abnormal beats, the PCC still exhibits a strong correlation of 0.860, with 15.0%~21.1% improvement compared with state-of-the-art approaches.

MADMFuse: A multi-attribute diffusion model to fuse infrared and visible images

In the field of deep learning vision, infrared and visible image fusion (IVIF) has received significant attention due to its ability to enhance scene comprehension by combining the complementary features of both types of images. Existing methods based on generative models are either unstable in training, such as generative adversarial network, or only have a single denoising object, such as diffusion models, resulting in poor fused results that cannot fully contain multi-modal features. To solve these problems, we propose a multi-attribute diffusion model to fuse infrared and visible images, termed MADMFuse. Specifically, to preserve the salient information in infrared images and the texture details in visible images, we have designed a forward diffusion process with shared noise for multiple independent attributes, ensuring that the denoising network learns complementary features of different attributes simultaneously during training. Subsequently, our reverse process with multi-attribute as conditions employs the denoising network that iteratively reparameterizes noise, gradually adjusting and achieving fine image fusion. Furthermore, to address the issue of feature degradation resulting from solely focusing on a single denoising object, we derive a multiple diffusion object loss function based on pixel-level fusion target, the fidelity and luminance term in this loss function aim to guide the fused results towards visual similarity with the source image while preserving its brightness. Extensive experiments indicate that MADMFuse is more effective than other state-of-the-art image fusion methods.

Efficient Photon Beam Diffusion for Directional Subsurface Scattering.

Real-time subsurface scattering techniques are widely used in translucent material rendering. Among advanced methods that rely on the bidirectional scattering-surface reflectance distribution function (BSSRDF), screen space algorithms exhibit limited translucency, while existing large-distance methods are inefficient and yield poor illumination details. To address these limitations for better large-distance scattering, we develop a novel algorithm by extending the photon beam diffusion (PBD) model within the light view and screen space. Unlike surface irradiance in prior methods, we incorporate the refracted beam in the medium into real-time scattering estimation, presenting a new consideration for photon beam utilization. Concretely, we store all photon beam samples in light view textures and utilize an adaptive sampling pattern for beam sample selection in large filtering kernel sizes. This can reduce the sample count based on surface attributes. In screen space, virtual sources are derived from samples to estimate PBD contributions, with an approximation that preserves boundary conditions. To avoid possible overestimation, we implement correction factors that scale contributions, effectively aligning our results with path-tracing references. Through these reformulations, our efficient PBD generates results closest to references among existing methods. The experiments accurately represent better front-face illumination details and backlit translucency effects, while significantly accelerating performance compared to previous large-distance methods.

Diffusion Models-Based Purification for Common Corruptions on Robust 3D Object Detection.

LiDAR sensors have been shown to generate data with various common corruptions, which seriously affect their applications in 3D vision tasks, particularly object detection. At the same time, it has been demonstrated that traditional defense strategies, including adversarial training, are prone to suffering from gradient confusion during training. Moreover, they can only improve their robustness against specific types of data corruption. In this work, we propose LiDARPure, which leverages the powerful generation ability of diffusion models to purify corruption in the LiDAR scene data. By dividing the entire scene into voxels to facilitate the processes of diffusion and reverse diffusion, LiDARPure overcomes challenges induced from adversarial training, such as sparse point clouds in large-scale LiDAR data and gradient confusion. In addition, we utilize the latent geometric features of a scene as a condition to assist the generation of diffusion models. Detailed experiments show that LiDARPure can effectively purify 19 common types of LiDAR data corruption. Further evaluation results demonstrate that it can improve the average precision of 3D object detectors to an extent of 20% in the face of data corruption, much higher than existing defence strategies.

Diffusion-based diverse audio captioning with retrieval-guided Langevin dynamics

Audio captioning, a comprehensive task of audio understanding, aims to provide a natural-language description of an audio clip. Beyond accuracy, diversity is also a critical requirement for this task. Human-produced captions possess rich variability due to the ambiguity of audio semantics (such as insects buzzing and electrical humming making similar sounds) and the existence of subjective judgments (metaphor, affections, etc.). However, current diverse audio captioning systems fail to produce captions with near-human diversity. Recently, diffusion models have demonstrated the potential to generate data with diversity while maintaining decent accuracy, yet they have not been explored in audio captioning. On the other hand, diffusion models tend to have low generation accuracy and fluency on text data. Directly applying diffusion models to audio captioning tasks may aggravate this problem due to the small size of annotated datasets and the mutable supervision target caused by the variability in human-produced captions. In this work, we propose a diffusion-based diverse audio captioning model by incorporating the BART language model into the diffusion model for better utilization of the pre-trained linguistic knowledge. We also propose a retrieval-guided Langevin dynamics module, which enables dynamic run-time alignment between generated captions and the target audio. Extensive experiments on standard audio captioning benchmark datasets (Clotho and AudioCaps) demonstrate that our model can achieve better performance on both accuracy and diversity metrics compared with state-of-the-art diverse audio captioning systems. The implementation is available at https://github.com/AI4MyBUPT/DACRLD.

Modelling white matter microstructure using diffusion OGSE MRI: Model and analysis choices

Alterations in white matter (WM) microstructure of the central nervous system have been shown to be pathophysiological presentations of various neurodegenerative disorders. Current methods for measuring such WM features require ex vivo tissue samples analyzed using electron microscopy. Magnetic Resonance Imaging (MRI) diffusion-weighted pulse sequences provide a non-invasive tool for estimating such microstructural features in vivo. The current project investigated the use of two methods of analysis, including the ROI-based (Region of Interest, RBA) and voxel-based analysis (VBA), as well as four mathematical models of WM microstructure, including the ActiveAx Frequency-Independent Extra-Axonal Diffusion (AAI), ActiveAx Frequency-Dependent Extra-Axonal Diffusion (AAD), AxCaliber Frequency-Independent Extra-Axonal Diffusion (ACI), and AxCaliber Frequency-Dependent Extra-Axonal Diffusion (ACD) models. Two mice samples imaged at 7 T and 15.2 T were analyzed. Both the AAI and AAD models provide a single value for each of the fit parameters, including mean effective axon diameter AxD¯, packing fraction fin, intra-cellular and Din and extra-cellular Dex diffusion coefficients, as well as the frequency dependence of Dex, βex for the AAD model. The ACI and ACD models provide this, in addition to a distribution of axon diameters for a chosen ROI. VBA extends this, providing a parameter value for each voxel within the selected ROI, at the cost of increased computational load and analysis time. Overall, RBA-ACD and VBA-AAD were found to be optimal for parameter fitting to physically relevant values in a reasonable time frame. A full comparison of each combination of RBA and VBA with AAI, AAD, ACI, and ACD is provided to give the reader sufficient information to make an informed decision of which model is best for their own experiments.

Simplifying bidisperse pore diffusion model for adsorbent felt

Mass transfer resistance within the adsorbent felt, governed by pore diffusion, limits adsorption efficiency. Current gas and solid side resistance models, which cover both micro- and macro-pore diffusion, are complex due to their numerous parameters. Sometimes, adsorbate transfer in the felt is predominantly controlled by macro-pore diffusion. This paper uses two key assumptions to introduce a simplified macro-pore diffusion model, distilled from a detailed macro–micro pore diffusion model. The model’s accuracy was confirmed by simulating two desiccant wheels, demonstrating its effectiveness in streamlining adsorbent felt analysis and predicting mass transfer dynamics.

Fast diffusion kinetics improves electron transfer regime selectivity to boost peroxymonosulfate activation

Electron transfer process (ETP) is barely influenced by half-life and migration distance in peroxymonosulfate (PMS) activation because it directly extracts electrons from pollutants for oxidation. However, the enhanced ETP selectivity is still challenging due to the poor mass transfer of PMS and pollutant. In this work, 2D porous N-doped carbon (PNC) with high reactant diffusion kinetics and multiple diffusion modes was designed to induce the interlayer confinement for efficient PMS activation with elevated ETP selectivity. Taking tetrabromobisphenol A as target pollutant, the confined PNC/PMS system achieved significantly enhanced PMS activation performance with the selectivity of ETP up to 97.8%. The vertical mass transfer of reactants into the inner interlayer space of the catalysts was proved to be promoted by the construction of molecule tunnels according to a diffusion model. Density functional theory calculations indicated that the subsequently triggered interlayer confinement facilitated the co-adsorption of PMS/pollutant on the interlayer surface and narrowed the energy gap of charge migration, resulting in enhanced PMS activation performance and ETP selectivity. Benefiting from the almost 100 % ETP selectivity, the confined PNC/PMS system exhibited substantially increased PMS utilization efficiency, admirable anti-interference ability towards various water matrices as well as long-term stability in a continuous-flow reactor. This work provides a new protocol for efficient and selective PMS activation by interlayer confinement in water remediation.

Differential diffusion modelling for LES of premixed and partially premixed flames with presumed FDF

Large eddy simulations (LES) with flamelet and presumed filtered density function closure are used to simulate turbulent premixed and partially premixed hydrogen flames. Different approaches to model differential diffusion are investigated and compared. In particular, two existing models are extended to the LES framework to correct the resolved diffusive flux of the controlling variables due to differential diffusion. A lean premixed turbulent hydrogen flame in a slotted burner configuration is simulated first to compare the capability of the considered models in capturing local mixture fraction redistribution, super-adiabatic temperatures and thermo-diffusive instabilities. Results show that both models describe the formation of cellular burning structures. Next, a partially premixed lifted hydrogen flame in vitiated hot coflow is simulated to gain insight on the relevance of differential diffusion modelling at a higher turbulence level, a different combustion mode and in the presence of a complex stabilisation mechanism. Good predictions of the turbulent mixing and temperature fields are observed. Moreover, results show that the flame lift-off height has an appreciable sensitivity to the differential diffusion model. When differential diffusion is included only in the thermochemistry database, only mild effects on the predicted temperature fields, mixing and flame height are observed. On the contrary, a considerable shift of the flame base is observed when corrections are applied in the LES at the resolved level, depending on what controlling variables are considered. Further analyses reveal how the corrections of diffusive fluxes in the thermochemistry and at the LES level affect differently the flame burning mode, whose details are given throughout the paper.

An investigation of the asphalt blending during recycling using the analytical solutions of the binder diffusion models

Diffusion occurs during the blending process of asphalt recycling. The coefficient of diffusion between any two diffusing materials in asphalt recycling is estimated as a constant value, which is found to vary in the order of 10−9–10−16 m 2 / s in past studies. In this study, Fick’s law of diffusion is solved analytically and modelled to simulate asphalt binder diffusion. A comparative study is performed among the models used to study the variability in asphalt binder diffusion. It is observed that the binder concentration gradients are inconsistent with similar initial and parametric values with the change in the boundary conditions adopted in the binder diffusion models. Similarly, the coefficient of diffusion predicted by the models varies for the models adopted under a constant time. From a comparative study of the models, it is observed that the complete blending condition may not be achieved in some cases.

Diverse Microstructures and Quasi-Ionic Liquid-like Transport Mechanisms in Concentrated "Water-in-Salt" Lithium Salt Electrolytes: A Molecular Dynamics Study.

"Water-in-salt"(WIS) electrolytes as potential green and nonflammable electrolytes are currently applied in various energy storage devices, such as lithium-ion batteries and supercapacitors. However, the microstructure at molecular scale and fast ion transport mechanism in such aqueous electrolytes are still under heavy debate due to the complex interactions among ions and water. Here, molecular dynamics simulations are used to study the microstructure and ion transport behaviors from the very dilute LiTFSI/water solution to the highly concentrated WIS electrolytes. It revealed that the diverse microstructures such as completely hydrated ions, ion complexes, and bridge-water molecules are jointly responsible for the electrochemical stability of WIS electrolytes. Diffusion model analysis showed that the Li+ ions exhibit a vehicular transport mechanism with first shell water molecules and structural diffusion mechanism with TFSI- anions. The lithium ion and its first hydration shell act as a single cationic entity. The entity forms a quasi-ionic liquid-like dynamic transport structure with associated anions. Our study challenges previous findings that the high transport dynamics of lithium ions arises from their transport in water-rich nanodomains in high-concentration WIS systems.

Study on the Diffusion Law of Grouting Slurry at the Pile Tip of Bored Piles in Gravel Pebble Layers

The post-construction improvement technology of bored pile tips has been widely used in the construction of gravel pebble layers, but the diffusion law of the grouting slurry remains to be revealed. To study the diffusion law of the grouting slurry at the tip of bored piles in gravel pebble layers, an equivalent relationship between the seepage continuity equation and the mass diffusion continuity equation was established based on an in-depth comparative analysis; further relying on the Diluted Species Transport Module in Porous Media of Comsol Multiphysics, a three-dimensional numerical model of slurry diffusion was established to systematically study the impact of key factors such as different porosities, grouting times, grouting pressures, slurry diffusion coefficients, and the ratio of vertical to horizontal slurry diffusion coefficients on the diffusion radius of the slurry. The calculation results show that with the continuation of grouting, the increase in porosity, grouting pressure, and slurry diffusion coefficient, or the decrease in the ratio of vertical to horizontal slurry diffusion coefficients, the diffusion range of the slurry increases accordingly. Furthermore, comparatively speaking, the impact of porosity is the smallest, while the impact of the slurry diffusion coefficient is the greatest. The research findings reflect the diffusion trend of the slurry in the gravel pebble layer, which has guiding significance for the quality control of the actual design and construction of grouting.

The developing hippocampus: Microstructural evolution through childhood and adolescence.

The hippocampus is a structure in the medial temporal lobe which serves multiple cognitive functions. While important, the development of the hippocampus in the formative period of childhood and adolescence has not been extensively investigated, with most contemporary research focusing on macrostructural measures of volume. Thus, there has been little research on the development of the micron-scale structures (i.e., microstructure) of the hippocampus, which engender its cognitive functions. The current study examined age-related changes of hippocampal microstructure using diffusion MRI data acquired with an ultra-strong gradient (300 mT/m) MRI scanner in a sample of children and adolescents (N=88; 8-19 years). Surface-based hippocampal modelling was combined with established microstructural approaches, such as Diffusion Tensor Imaging (DTI) and Neurite Orientation Dispersion Density Imaging (NODDI), and a more advanced gray matter diffusion model Soma And Neurite Density Imaging (SANDI). No significant changes in macrostructural measures (volume, gyrification, and thickness) were found between 8-19 years, while significant changes in microstructure measures related to neurites (from NODDI and SANDI), soma (from SANDI), and mean diffusivity (from DTI) were found. In particular, there was a significant increase across age in neurite MR signal fraction and a significant decrease in extracellular MR signal fraction and mean diffusivity across the hippocampal subfields and long-axis. A significant negative correlation between age and MR apparent soma radius was found in the subiculum and CA1 throughout the anterior and body of the hippocampus. Further surface-based analyses uncovered variability in age-related microstructural changes between the subfields and long-axis, which may reflect ostensible developmental differences along these two axes. Finally, correlation of hippocampal surfaces representing age-related changes of microstructure with maps derived from histology allowed for postulation of the potential underlying microstructure that diffusion changes across age may be capturing. Overall, distinct neurite and soma developmental profiles in the human hippocampus during late childhood and adolescence are reported for the first time.

Conditional guided generative diffusion for particle accelerator beam diagnostics

Advanced accelerator-based light sources such as free electron lasers (FEL) accelerate highly relativistic electron beams to generate incredibly short (10s of femtoseconds) coherent flashes of light for dynamic imaging, whose brightness exceeds that of traditional synchrotron-based light sources by orders of magnitude. FEL operation requires precise control of the shape and energy of the extremely short electron bunches whose characteristics directly translate into the properties of the produced light. Control of short intense beams is difficult due to beam characteristics drifting with time and complex collective effects such as space charge and coherent synchrotron radiation. Detailed diagnostics of beam properties are therefore essential for precise beam control. Such measurements typically rely on a destructive approach based on a combination of a transverse deflecting resonant cavity followed by a dipole magnet in order to measure a beam’s 2D time vs energy longitudinal phase-space distribution. In this paper, we develop a non-invasive virtual diagnostic of an electron beam’s longitudinal phase space at megapixel resolution (1024 × 1024) based on a generative conditional diffusion model. We demonstrate the model’s generative ability on experimental data from the European X-ray FEL.