<p>Seamounts provide valuable insights into plate tectonics, volcanic activity, and the geological evolution of the oceanic plates. Their shapes, sizes, and distributions are crucial for reconstructing past seafloor spreading and hotspot activity. Accurately detecting the shape and location of global seamounts on bathymetric maps has long been a challenging task, whereas deep learning offers a highly feasible solution for this endeavor. This study approaches the global seamount distribution analysis as a semantic segmentation task, for which we manually annotate a seamount semantic segmentation dataset on bathymetric data. We trained an improved Segformer model specifically designed for this task, which achieved significant improvements over the original Segformer and other models, with a mean Intersection over Union (mIoU) of 84.97% and a Dice coefficient of 94.12%. By applying the proposed model to predict seamounts on a global bathymetric data of 15-arcsecond resolution, we obtained a detailed spatial distribution of 51,965 seamounts globally and further analyzed these predictions statistically. The results reveal that seamounts in the Pacific Ocean dominate globally, accounting for 61.05% of the total area and 60.79% of the total volume. The proportions in the Atlantic Ocean and Indian Oceans are comparable, while the contribution from the polar oceans is less than 3%. Additionally, the summit depth of seamounts significantly deepens with increasing plate age within 150 Ma. This new dataset of global seamounts with accurate shapes provides a valuable tool for further investigating mechanism of seafloor volcanism and seamount formation.</p>