|
Squeeze-and-Excitation: To recalibrate feature maps and improve the model's emphasis on more crucial channels, the Squeeze-and-Excitation (SE) module adds a channel-wise attention mechanism, as shown in Figure.
Squeeze and excitation are their two primary functions. Global Average Pooling compiles spatial data into a channel description during the squeeze phase.
Subsequently, the excitation phase generates channel significance weights using two Dense layers: sigmoid activation and ReLU activation to minimize dimensions.
These weights are multiplied by the input feature maps to emphasize important channels and suppress less important ones. By emphasizing important
|