Sonar imaging provides representative real-time mapping of the water bottom, which is particularly valuable in search-and-recovery operations, as it helps quickly locate drowned victims. However, practical analysis of sonar data is often challenging due to its complexity and automated recognition demanding substantial training data. In this paper, we present an extension to SDVD (Sonar Drowned Victims Dataset), a sonar image dataset with annotated binary masks marking drowned victims for training segmentation models. To improve segmentation processing, we propose Multi-Channel Input Augmentation (MCIA). This method augments the segmented image with additional channels, increasing it from one to seven, thereby enhancing the model’s information input. To achieve this, each channel represents a processed original image by applying: Sobel edge detection, Haar wavelet, High-Pass features, pixelated map, CLAHE and distance transform. To evaluate segmentation performance, we tested different models on all SDVD subsets, alongside various data augmentation techniques, such as RICAP. Among the models, MCIA-U 2 Net demonstrated the highest accuracy, achieving an Intersection over Union (IoU) of 75.89% and a Dice Score of 86.30% on base SDVD, 71.57% IoU and 83.43% Dice Score on CleanedSDVD, and 68.55% IoU and 81.33% Dice Score on ExtendedSDVD. Additionally, we utilized a channel-level Local Interpretable Model-Agnostic Explanation (LIME) method for detailed evaluation underlining the efficiency of the proposal. These results highlight the effectiveness of the proposed approach in training segmentation models and indicate significant potential for further enhancements in sonar image recognition methods for victim recovery applications.

Federated learning (FL) has emerged as a practical solution for training deep neural networks while preserving data privacy, particularly in scenarios like the Internet of Things (IoT), where user devices generate sensitive private data. While FL addresses privacy concerns, existing aggregation methods remain vulnerable to malicious poisoning attacks. Therefore, developing new approaches to optimize client selection as a preventive mechanism is crucial to FL’s strategy. This paper introduces a novel enhancement to FL by integrating the MultiMOORA technique into a robust aggregation mechanism with knowledge distillation (KD). During FL rounds, local models are evaluated using classification metrics (loss, accuracy, precision, recall, f1-score) and ranked using the Multi-Criteria Decision Maker (MCDM). The resulting ranking is used to select the best-performing models for aggregation and detect possible infected clients. Furthermore, the top client is chosen as an aggregator. During aggregation, chosen best models are weight-averaged and form teacher ensemble, used for stabilizing newly formed global model via KD. The proposed FL strategy reduces the risk of adversarial poisoning attacks on FL systems and stabilizes FL training process. The experimental results indicate the system’s resilience and ability to perform reliably in hostile environments, as the proposed FL system achieved 89.50% F1-score under label-poisoning scenario. Based on further analysis, the proposed methods could be integrated into a digital-twin-based modern healthcare system.

Decentralized tasks use machine learning models with certain assumptions. The first is the sharing of weights and feature extractors. The second is maintaining the privacy of the data. The idea of learning using multiple models can also be applied in parallel training, where a given model is trained on a different thread. This has applications in creating models based on federated learning, the Internet of Things and Digital Twins. This paper proposes a new neural network model that uses the atrous technique and attention modules. In addition, we propose a hierarchical-based training strategy, where the best model shares weights and is omitted during further training. This reduces the number of training epochs and increases the model's adaptability to a given set. The tests conducted on a publicly available medical database indicate high learning potential for both the proposed model and the hierarchical learning strategy.

Agricultural health monitoring is a critical task in ensuring the stability of modern agriculture. Many plant diseases share visual similarities, making manual inspection both time consuming and error prone, which is why robust and adaptable disease detection frameworks are not only desirable but essential to maintaining a resilient agricultural ecosystem. In this paper, we propose a hybrid federated learning (FL) framework that integrates a globally shared feature extractor with a client-specific self-attentive branch and classifier. The proposed framework uses a global model with both globally shared and client-tailored branches to achieve better performance for specialized tasks in decentralized training scenarios. The experiments were carried out on a Plant Village data set in a scenario, where each client represented a different crop type and faced a different leaf disease classification problem. The proposed solution revolved around the clients sharing the global weights, thus simultaneously contributing towards better feature extraction of the common leaf features, while the specialized segment of the model focused on proper interpretation of the extracted features (via cross-attention mechanism) and direct classification. The results obtained demonstrate the effectiveness of the proposed approach over standard local training, as training with the proposed hybrid FL framework resulted in a perfect classification of the precision 100% of apple leaf disease.

3D point clouds (LiDAR) are an integral application in robotics, creating 2D or 3D maps that help autonomous vehicles navigate or avoid and recognize obstacles as they drive. In this work, we propose Graph Convolutional Network with Multi-Scale Pooling to classify three-dimensional objects. Previous research very rarely undertakes the transformation of point clouds into graphs and the use of increasingly popular graph networks. Researchers focus on transforming 3D points into voxels and using classical CNN or 3D CNN convolutional networks or methods that directly use the entire cloud, as is evident in the PointNet model. Therefore, we focused on transforming our data into irregular structures (graphs) and classifying them. Our proposed method increases the possibilities for point cloud interpretation by taking into account the relationships between points in space. In order to achieve even better accuracy results, we use the integration of two pooling techniques in the final stage of the model. Our GCN with Multi-Scale Pooling architecture has been trained and tested on the ModeiNetl0 dataset, achieving high-quality metrics: Accuracy, Precision, Recall, and F1-Score compared to other classical SOTA models in the classification domain. Additionally, we carried out experiments based on ShapeNet Core dataset, for which our network achieved almost 100% accuracy. In addition, we showed that our proposed model is built from a much smaller number of parameters compared to other modern methods.

Digital twins are a technology that allows for a virtual copy of a real object or process. The main goal is to map specific features to prevent problems or monitor conditions and operations. In this work, we propose a framework for a sonar system. Data acquired by sonar are most often sent to an operator or classification network to detect objects on the seabed. However, such a classifier needs a large amount of data to be properly trained. Therefore, we propose a digital twin model that uses generative adversarial networks (GANs) with feature fusion to obtain synthetic data for further processing. The proposed GAN model is based on dual generators with combining results that are passed further. The proposed technique indicates that building an artificial intelligence module by fusing real and synthetic data is important and allows for achieving high augmentation results in sonar applications.

Taking photos against the light generates a certain visual effect. Taking a photo in the direction where the sun is located also results in a change in the temperature of the photo as well as the appearance of a visual effect in the form of the flare of light. In this article, we present an innovative neural network model called Y-ResNet, whose input consists of two samples and the output consists of one. This solution makes it possible to train the network by providing the original image and the flare effect, which will result in a modified sample. The training was conducted on a commonly known CityScapes dataset, where, by using classic data processing methods and the k-means algorithm, it was possible to add a flare if there was a visible portion of the sky in the input image. The proposed solution was described and tested to demonstrate the capabilities of the proposed method. The results show the superiority of the approach against the traditional ResNet without a second encoding path, generating better results, and creating a better impression of the lens-flare effect.

Analysis of the seabed using sonar is a key technology enabling the assessment of the substrate, detection and classification of objects located there. However, quite often sonar data is processed by users due to the small amount of measurement data. This is due to the need to create large data sets, and creating a sonar image is often dependent on atmospheric conditions. In this paper, we present a solution based on digital twins that allows the implementation of a digital twin layer for sonar applications. A digital twin layer based on generative and classification network models increases the amount of data and improves the effectiveness of solutions. For this purpose, we propose multi-attention models that focus on local and global sonar features and enable their fusion. Moreover, a technique for exchanging weights between networks in such a solution was modeled to reduce the amount of computing power. The proposed approach allows for analyzing images by focusing on different features and increasing the automatization of processing its data. To verify the operation, various sonar data were used and high classification accuracy was achieved as well as the generation of new data.

Autonomous vehicles are a key element of the automotive industry, where the impact of the human factor on the condition of the vehicle and driving is minimized. An important element is the analysis of vehicular condition, which allows maintainence of its value and correct operation. We propose a system based on the analysis of the image of vehicles, which determines whether there is any damage. For this purpose, we propose a new model of a Convolutional Neural Network (CNN) that has 0. 395M trained values. The architecture of the network is adapted to the analysis of spatial features that allow networks to be adapted to analyze primarily vehicular shape and orientation in relation to other objects. The model also implements spatial dropout and regularization techniques for preventing overtraining and maintaining model generalization. The modeled architecture contributes to obtaining high classification accuracy at 94.78% using a public database and exceeding metrics of known transfer learning models.