Identity-preserving portrait generation has achieved tremendous advancements with the development of diffusion models. However, multi-ID generation remains challenging due to degraded identity fidelity and insufficient control over layout, pose, and expression. To address these challenges, we propose GroupPortrait, a novel approach for multi-ID portrait generation with three key innovations:(1) LatentID for high-fidelity identity preservation, (2) Facial Controller enabling layout guidance and fine-grained facial control, and (3) Mask-Attention Controller allocating identity embeddings to specific facial regions. First, the LatentID module improves identity preservation by adding LatentID loss during training. It maps latent representations to identity features and uses ID consistency loss for feedback training to improve identity retention. Since LatentID loss is calculated in latent space, it is more efficient in terms of time and GPU usage compared to the method that calculates ID loss in pixel space. Second, to enhance layout and facial controllability, the Facial Controller utilizes 3D Morphable Models (3DMM) to acquire facial shapes, poses, and expressions for each individual, imposing strong spatial conditions during the diffusion process. Finally, we propose a novel Mask-Attention Controller for multi-ID generation, which distributes ID embeddings into target facial regions by aligning the cross-attention map of LatentID with the given facial region masks. Extensive experiments demonstrate that GroupPortrait can generate group portraits with high fidelity, local harmony, and controllability.

Ground Penetrating Radar (GPR) is widely used in geotechnical engineering investigations, construction quality assessment, and geological disaster surveys due to its high resolution, accuracy, and non-destructive testing capabilities. However, the accuracy of GPR inversion imaging is often compromised by climatic conditions (such as precipitation and temperature) and complex subsurface environments, leading to suboptimal performance. To address this issue, we propose a Shuffle Attention U-Net Generative Adversarial Network for GPR inversion imaging—SAU-GAN. This network consists of a generator and a discriminator. The generator features an encoder-decoder network enhanced with a Shuffle Attention mechanism, facilitating efficient feature extraction from B-scan images and aiding in the generation of permittivity models. The discriminator evaluates generated models against real ones, providing feedback to supervise the generator’s performance. Both components use double normalization to stabilize parameters and convolutional outputs. Additionally, a multi-scale structural similarity (MS-SSIM) loss function enhances the existing loss function, significantly improving inversion results. Experiments with synthetic data demonstrate that SAU-GAN produces permittivity models with higher accuracy and clearer boundaries than existing methods. Even under interference, it is able to perform precise inversion, demonstrating outstanding robustness and generalization performance. We conduct a quantitative analysis of SAU-GAN using SSIM, PSNR and MSE metrics, further validating its superior performance. When applied to real measured data, SAU-GAN also exhibits commendable performance, validating its effectiveness and practical value.

Traditional ground-penetrating radar (GPR) inversion techniques, while capable of providing high-resolution subsurface imaging, suffer from issues, such as heavy reliance on initial models, high computational demands, and sensitivity to noise and data incompleteness. In contrast, deep-learning-based methods excel in feature extraction and model fitting. However, as a data-driven algorithm, the practical application of convolutional neural networks (CNNs) is limited by the quantity of labeled samples. To reduce the dependence of CNN-based GPR inversion methods on observational data and labels, this project proposes an inversion method based on closed-loop CNNs (CL-CNNs). This approach improves inversion accuracy and reduces the ill-posedness of GPR inversion by modeling both the forward and inverse GPR processes. The CL structure increases the number of features that CNNs can learn from limited labeled samples, while the mutual inversion constraints between the forward and inverse subnetworks help alleviate the ill-posedness of the inversion problem, making the inversion results more consistent with geological principles. Research using synthetic data demonstrates that this method outperforms traditional approaches, as evidenced by enhanced structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR), and a significantly lower mean-squared error (mse), highlighting its advanced performance compared with traditional open-loop CNNs (OL-CNNs). Furthermore, applying this method to real measurement data further validates its effectiveness and practical applicability in engineering contexts, emphasizing its significant practical value.

Electrical resistivity tomography (ERT) is a key geophysical technique that provides detailed information on subsurface structures by measuring the distribution of electrical resistivity underground. ERT suffers from limitations in electrode arrangement, interference from environmental and instrument noise, and existing data processing algorithms that fail to adequately consider geological heterogeneity and uncertainty, resulting in insufficient inversion resolution. Traditional ERT methods rely on simplified algorithms and a limited number of observation points, which smooths model details and further reduces resolution. To address the resolution issues in ERT, this article proposes a deep learning inversion method that integrates prior physical information. This method uses low-resolution inversion results as prior knowledge to provide the deep learning algorithm with a constrained initial model, thereby combining the physical basis of traditional methods with the data-driven advantages of deep learning. The method not only retains the strengths of traditional inversion but also enhances the resolution and imaging efficiency of the inversion model using deep learning technology. Synthetic data experiments demonstrate that integrating deep learning significantly improves the model’s ability to detail subsurface structures, especially in the transition zones of shallow structures and the recovery of deep anomalies. Results from measured data indicate that the proposed method not only achieves high-resolution inversion but also maintains good consistency with prior information.

Ground-penetrating radar (GPR) inversion typically relies on iterative methods, which often involve high computational complexity and challenges in noise handling. These limitations affect the robustness and generalization of traditional approaches. To address these issues, we propose an innovative inversion method using diffusion models (DGPRI-Net) tailored for GPR. Diffusion models inherently capture signal characteristics through progressive noise addition and subtraction, reducing noise impact and enhancing robustness. This approach effectively overcomes the noise management weaknesses of conventional methods. In the reverse generation process, we use the UNet++ network architecture, enhanced with vision transformer (ViT) structures and a simple parameter-free attention module (SimAM). This combination improves multiscale feature extraction and contextual understanding, increasing robustness and enabling high-precision permittivity models. To further evaluate the robustness of the model, we prepared three dedicated test sets: one with added noise, one without low-frequency signals, and one with 30% of the columns missing. Comparative experiments with synthetic data showed exceptional inversion accuracy, superior noise management, and enhanced robustness and generalization. We also validated performance with metrics such as structural similarity index measure (SSIM), peak signal-to-noise ratio (PSNR), and mean squared error (mse). Applied to measured data, our method continued to yield impressive results, confirming its practical value and effectiveness. This study highlights the potential of diffusion models in advancing GPR inversion applications.

Although deep learning techniques for Ground Penetrating Radar (GPR) inversion offer significant advantages, such as high accuracy and computational efficiency, they still face challenges related to limited robustness and generalization. Using envelope radar data in inversion helps mitigate some of these issues. This data emphasizes amplitude characteristics, reducing the impact of high-frequency noise and phase-related problems on the inversion results. In this paper, we propose a GPR inversion method based on a conditional generative adversarial network (cGAN), incorporating envelope radar data as conditional input for both the generator and discriminator. We also use double normalization to adjust the discriminator’s convergence speed, ensuring the adversarial relationship is maintained, which enhances model robustness without sacrificing inversion accuracy. For the loss function, we introduce a hybrid of Mean Squared Error (MSE) and Multi-Scale Structural Similarity Index Measure (MS-SSIM) loss, guiding the generator to produce results closer to the true model. To evaluate the inversion performance of our proposed method, we designed three synthetic data experiments: a comparison experiments with varying degrees of low-frequency component depletion, a comparison experiments with different noise levels and a comparison experiments with varying central frequencies. Results show high resolution, clear inversion boundaries, and well-defined anomalous structures, demonstrating high inversion accuracy and robustness. Furthermore, our method shows superior performance and practical value with real-world data.

Magnetic anomaly pole transformation converts magnetic field data into an equivalent response at the true magnetic pole, eliminating shifts and distortions, simplifying the interpretation of subsurface magnetic bodies, and improving data interpretation and inversion accuracy. However, the main challenge in magnetic anomaly pole transformation lies in the nonlinear nature of the signals, making traditional methods difficult to apply. The interaction between the shape, depth, and magnetic inclination of magnetic bodies, especially in high- and low-latitude regions, can distort the transformed signal, leading to unclear causal relationships. To address this, this article proposes a deep learning-based approach that automatically extracts high-dimensional features and establishes nonlinear mappings to enhance the correlation between magnetic anomaly signals and geological structures. Deep learning does not require explicit physical models and, through training with large datasets, demonstrates stronger robustness and accuracy, especially in areas where traditional methods fail. The proposed method is validated using both synthetic and measured data. Synthetic data simulates magnetic bodies of various shapes, depths, and magnetic inclinations, confirming the method’s stability and accuracy in handling complex nonlinear signals. The measured data evaluates its pole transformation advantages in typical ore deposit regions. The results indicate that the deep learning model significantly enhances the accuracy of pole transformation, particularly in areas with complex magnetic anomaly signals, effectively preventing signal distortion and demonstrating exceptional generalization capabilities.

Certain ore deposits feature both density and electrical anomalies, making them detectable via gravity and electromagnetic (EM) methods. However, under complex field conditions, signals are often distorted or lost due to observational errors, undermining inversion reliability. In joint inversion, errors from a single data source may mislead the overall model, resulting in structural deviations and blurred orebody boundaries. Additionally, gravity and EM inversions exhibit different volume effects, often causing inconsistencies in spatial scale representation. Their distinct physical mechanisms further hinder the establishment of clear nonlinear mappings, limiting the effectiveness of traditional joint inversion approaches in achieving consistent integration and stable results. To address these challenges, we propose a Spatial Density-Informed Electromagnetic Inversion Network (SDI-EMI Network), a deep inversion network that fuses spatial density and EM response data. The network first performs gravity inversion to estimate orebody geometry, which serves as a structural prior for guiding EM inversion. By aligning volume deformation patterns, it ensures unified scale representation, enhances data complementarity, and suppresses interference from non-orebody regions. This method overcomes limitations in prior modeling and data fusion while leveraging deep learning’s nonlinear capacity. Experimental results confirm that SDI-EMI Network. offers improved resolution, structural clarity, and robustness for identifying deposits with coexisting density and resistivity anomalies, supporting its potential in complex geological settings.