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.

Analysis of the immediate environment and its segmentation is particularly useful in the mechanism of autonomous vehicles. Accurate marking of important points and places while traversing the route allows cars to analyze the space and navigate safely. Transformers models and their successors such as UNETR and Vision Transformer using Multi-Head Attention are widespread in the segmentation process. These approaches permit to better image classification and segmentation results, as each head in the transformer simultaneously analyzes different aspects of the relationship between image fragments (patches). Given the high potential of transformer algorithms, in this article we present an extended UNETR model, which in its design additionally includes the RSAM attention block (Residual Spatial Attention Module) for the road segmentation process. Our RSAM-UNETR architecture supported by additional weights in the loss function performed perfectly and achieved Accuracy = 99.04%, mDice = 89.21% and mIoU = 81.80% with a very stable training process achieving a higher mDice value by 5.46% and mIoU by 7.31% compared to the classic UNETR.