Search

To solve the problems of higher model complexity and difficulty in constructing regularization items in the semi-supervised classification model, a new semi-supervised image classification model named AGSH from the perspective of enriching sample feature representation is constructed, which is fusion with an adaptive graph structure. The model AGSH introduces the adaptive graph convolutional neural network AGCN, aiming to extract the relationship between the features of the CNN model based on the convolutional neural network model CNN. The analysis of the generalization performance of the AGSH model also shows the effectiveness of solving semi-supervised related problems. The experimental results show that the accuracy of the AGSH model is improved compared with that of the single CNN model on the five image datasets. The research expands the content of the semi-supervised image classification algorithm and provides an essential reference for the modeling method to solve the few-sample classification problem. [ABSTRACT FROM AUTHOR]

The analysis dictionary learning has gradually been applied to various classification tasks in recent years. However, most of the existing classification methods based on analysis dictionary learning are only pursuing simpler optimization in the training phase, higher speed in the testing phase and paying less attention on the discriminative research of the dictionary. In response to the above problems, this paper constructs a two-layer structure and super vised task-driven analysis dictionary learning (Task-driven analysis dictionary learning, TADL) model. The structured analysis dictionary in the first layer structure improves the discriminability of the TADI model. The second layer structure uses the trace quotient criterion to reduce the intra-class divergence, increase the inter-class divergence, and transform the classification problem into a low-rank constraint discriminant analysis dictionary problem. From a global perspective, the robustness of the model has been enhanced. Finally, by comparing with two state-of-the-art classification methods, the advantages of the TADL model are verified. [ABSTRACT FROM AUTHOR]

The storage of massive point cloud data has great significance to the real-time 3D collaborative perception of autonomous driving. However, due to the requirements of data security and confidentiality, some data owners are unwilling to share their private point cloud data, which limits the improvement of model training accuracy. Federated learning is a computing paradigm that focuses on data privacy and security. This paper proposed a novel approach based on federated learning to address the challenge of large-scale point cloud semantic segmentation in collaborative vehicle perception scenarios. It integrated position encoding with inter-point angle information and geometric diffraction of neighboring points to enhance the feature extraction capability of the model. Finally, it dynamically adjusted the aggregation weights of the global model according to the generation quality of the local model to improve the ability to maintain the local geometric structure of the data. This paper applied the proposed method on three datasets, such as Semantic KITTI, SemanticPOSS and Toronto3D. The results show that the proposed approach significantly outperforms the single training data and the FedAvg-based method, and fully exploits the value of the point cloud data while taking into account the privacy sensitivity of each party's data. [ABSTRACT FROM AUTHOR]

Most existing class incremental learning methods employ data storage or extended network structures, but they cannot effectively alleviate the catastrophic forgetting problem due to memory resource limitation. To solve this issue, this paper proposed a brain-inspired generative replay approach. Firstly, it used VAE-ACGAN to simulate the memory self-organizing system to improve the quality of the generated pseudo-samples. Then, it used a shared parameter module and a private parameter module to protect the extracted features. Finally, the potential variables in the generator used a Gaussian mixture model to select specific replay pseudo-samples. The experimental results of the proposed method on MNIST, Permuted MNIST and CIFAR-10 show that the classification accuracy of the proposed method is 92.91%, 91.44% and 40.58%, respectively, which is significantly better than other class incremental learning methods. Furthermore, the backward transfer and forward transfer metrics reach 3.32% and 0.83% on the MNIST dataset, demonstrating that the model achieves a trade-off between task stability and plasticity, effectively preventing catastrophic forgetting. [ABSTRACT FROM AUTHOR]