This study presents a dance movement style transfer method based on 3D pose estimation and spatiotemporal feature fusion. The proposed method aims to preserve the structural characteristics and temporal coherence of source dance movements while adapting them to a target style. To address the challenges of movement fidelity, style consistency, and motion smoothness, the framework employs a Manifold Event Forecaster as the core module. This module integrates a Counterfactual Manifold Optimizer, Modeling Discrete Dance Events, and a Comprehensive Framework for Dance Motion Transformation to predict and transform dance motion sequences in a style-aware manner. The Counterfactual Manifold Optimizer guides the generation of target-style movement trajectories in the motion manifold, while the discrete event modeling module segments key dance events for more precise motion transformation. Stochastic refinement is introduced to improve the robustness of style transfer under motion variability and pose uncertainty. Experimental evaluations show that the proposed method can generate coherent target-style dance movements while maintaining the choreographic structure and motion characteristics of the original sequence. The study provides a feasible framework for dance movement style transfer and offers potential applications in digital choreography, motion synthesis, and creative performance analysis.
Geological disasters threaten ecological stability, yet restoration planning in high-risk areas remains difficult because deformation, environmental conditions, and hazard drivers vary across space and time. We developed a multimodal dataset, HGR-InSAR-2022, covering 12[Formula: see text]640[Formula: see text]km 2 and comprising 312 Sentinel-1 scenes, 1486 Sentinel-2 images, a 30-m digital elevation model, rainfall and geological layers, and 1276 documented landslide locations collected from 2018 to 2022. The dataset provides comprehensive spatial and temporal observations for characterizing terrain instability, ecological degradation, and heterogeneous environmental responses across disaster-prone landscapes. Based on this dataset, we propose AGRNet, an adaptive geospatial restoration network that integrates a constraint-aware restoration optimizer, an agent-based risk mapping unit, and an uncertainty-driven InSAR fusion module. The network combines InSAR-derived deformation, multispectral ecological indicators, topographic variables, and hazard factors to generate ecological security patterns and restoration-priority maps while quantifying prediction uncertainty. This multimodal integration enables the model to capture complex interactions between surface deformation, ecological conditions, and geological hazard processes at different spatial scales. On an independent 20% test set, AGRNet achieved an [Formula: see text]1-score of 0.91 and an intersection-over-union of 0.84 for ecological security classification, outperforming random forest and conventional convolutional neural network baselines by 8.7 and 5.1 percentage points, respectively. Its restoration-priority map achieved an area under the receiver operating characteristic curve of 0.94, while uncertainty modeling reduced deformation prediction RMSE by 16.3%. These results demonstrate that the proposed framework provides accurate, uncertainty-aware support for ecological restoration and geological-disaster mitigation.
Ultrasound imaging is an important non-invasive technique to evaluate chronic infections like hepatitis, but the diagnostic accuracy of the method is low due to speckle noise, poor resolution, and the lack of labeled clinical data. Conventional deep learning models have been found to perform well, but are usually limited by the lack of adequate training samples and inefficient processing of noise and texture in ultrasound. To address these problems, this paper suggests a new three-step pipeline combining synthetic data generation, denoising, and super-resolution with subsequent validation using a baseline classification model. An Adaptive Conditional GAN with diffusion modeling is used to generate a synthetic dataset for the hepatitis affected liver ultrasound images with pathological variability. A Multi-Scale Attention CycleGAN denoiser is used to remove speckle noise while preserving both texture and fidelity. Along with that, the Residual Dense Transformer Super-Resolution module was used to boost structural and fine-grained information of the ultrasound images by fusing the cross-scale features. Then, several deep learning classifiers such as ResNet, DenseNet, EfficientNet, and Hybrid_CNN (ResNet-DenseNet) were employed to evaluate the diagnostic utility of the reconstructed images. Extensive experiments revealed that the proposed framework (Epoch 500) performed better than state-of-the-art approaches in PSNR (28 dB), SSIM (0.75), MSE (0.010), MAE (0.020), and LPIPS (0.150) of image reconstruction. The EfficientNet classifier achieved an overall average classification accuracy of 90% at Epoch 500. These findings suggest that the proposed model produces high-fidelity ultrasound images and empowers trustworthy staging of hepatitis.
The process of verifying or identifying individuals based on the combination of two or more biometric modalities is known as multimodal biometric authentication. Multimodal biometric authentication has gained significant attention for enhancing the security of biometrics. Several techniques were developed to perform efficient multimodal biometric authentication. However, previous techniques faced issues in terms of low performance, high error rates, and overfitting problems while handling noisy and imbalanced biometric data. To overcome these issues, a novel optimized feature fusion-based Deep Learning (DL) model for multimodal biometric authentication was developed. The key objective of the proposed framework is to perform multimodal biometric authentication by combining iris, fingerprint, and signature biometrics. The proposed model improves authentication accuracy and reliability by reducing error rates, overfitting, and generalizability across varied biometric modalities. Initially, normalization is used to normalize the image, and Weighted Gaussian Filtering (WGF) is used to reduce the noise. Then, augmentation techniques are applied to solve the imbalance in classes. A Spatial Attention-aided EfficientNet-B0 (SA-EB0) model is utilized to extract essential features by focusing on the most informative sections of each biometric input. A Feature Fusion-based Extended Convolutional Ghost Network (FF-ECGN) is used to fuse the gathered features and also to identify authenticated and unauthenticated data. The feature fusion integrates complementary information obtained from iris, fingerprint, as well as signature to enhance the reliability of authentication. The Chaotic Artificial Humming Bird (CAHB) algorithm is applied to fine-tune model parameters, hence improving the model’s performance. The model’s performance was evaluated through accuracy, precision, recall, [Formula: see text]1-score, specificity, and error metrics MAE, MSE, and RMSE. The proposed approach performs better with 98.55% accuracy and a much lower error rate of 0.0107 EER compared with other existing models.
Waste Classification and management are crucial to improving the environment, with the ultimate goal of promoting economic recycling and environmental safety. Due to the vast diversity of waste, relying on manual Waste Classification and recycling waste products remains ineffective and costly. Advancements in Computer vision and Deep Learning technologies have been applied for Intelligent waste classification. However, the existing methods are incompetent to capture the discriminative features of waste from different classes and struggle with handling complex backgrounds. Consequently, this research proposes the Virions coordinated colony search optimization-based conditional learning with pooling enhanced convolutional neural network (VC2S-CPEnCN) to improve waste classification. Specifically, the proposed method exploits the Dense blockwise directional pattern (DeBDP) for extracting intrinsic features that provide better characterization of granular pixel regions. Further, the incorporation of conditional learning dynamically adapts the model to learn the complex relationships of waste objects even in varying lighting, occlusion, and background clutter conditions. In the proposed model, the application of T-max average pooling promotes adaptive feature retention without adding computational overhead. In addition, the hyperparameters are optimally tuned using the Virions coordinated colony search optimization (VC2S) that promotes faster convergence and improves the classification accuracy. Experimental results of the VC2S-CPEnCN are assessed using various metrics, exhibiting 96.78% accuracy, 96.83% F1-score, 96.97% precision, and 96.68% recall in the Garbage Classification dataset, outperforming the other conventional methods.
