Attention calculation process in the Swin Transformer block. The left panel represents the use of W-MSA in the layer l, the middle panel represents the use of SW-MSA in the next layer, and these two appear in pairs. The attention calculation in SW-MSA is shown in the right panel, where the completion of the new window is achieved by moving the windows and then a masking operation is applied to calculate attention within the window. Please note that the black squares represent windows with a size of 4 × 4, and the grey squares represent patches.