A Dynamic Multi-view Fusion Model of Transmission Lines Icing Classification

In this paper, we propose a novel dynamical fusion model, named Meta-Multi-View-Learning (MMVL), built on the large language model and meta-learning approach. The model takes the multi-view learning of each instance as an independent task and fuses the multi-view signal according to their content. Specifically, we devise a Transformer-based multi-view meta encoder to learn the meta information by modeling the correlations between visual information from different views. By utilizing the metainformation, we generate fusion function parameters for each instance and dynamically fuse multi-view signals for representation learning. Ultimately, by incorporating the representations with a pre-defined prompt pattern, we can predict the icing types (i.e., glaze, rime, hoarfrost, snow, and non-icing cover) empowered by the knowledge in a pre-trained large language model. Through extensive experiments on three datasets, TL-Icing, Fashion-MV, and Caltech101-20, we demonstrate that our proposed model is able to significantly outperform state-of-the-art methods.