t-SNE Visualization: Unveiling Discriminative Features in High-Dimensional Data

t-SNE is commonly employed using nonlinear techniques to visualize high-dimensional data in a low-dimensional format. In our study, we utilize t-SNE to present the results of visualizing linear features on the final layer of the model. We perform feature visualization on both the backbone and the proposed model. To enhance the visualization, we randomly select 30 categories from the CUB-200-2011 dataset for the experiment. From Fig.6 to Fig.8, it is evident that the proposed method learns more discriminative visual representations. In Fig.6, the benchmark network TinyVit successfully groups most similar samples into an aggregated region, although there are still some samples that aggregate in more distant regions. In Fig.7, the feature augmented with only the SFRL module is able to group all similarities into a general region, but some regions exhibit confusion between categories. In Fig.8, we combine the SRFL and OL modules. The high-dimensional features output by the overall model significantly reduce cross-regions between different categories after undergoing low-dimensional processing.