The Method of Portrait Segmentation Based on Efficient Channel Attention Mechanism

The attention mechanism can extract task-relevant vital information while suppressing less important information, playing an increasingly critical role in deep feature representation for semantic segmentation. This research proposes a neural network model based on the widely used U-Net architecture, incorporating the ECA (Efficient Channel Attention) attention mechanism. To address the issue of edge segmentation blur, an improved module based on ASPP (Atrous Spatial Pyramid Pooling) is introduced into both the compression and expansion paths of U-Net. Additionally, to tackle the issue of neglected global context information at different scales in portrait segmentation networks, a multi-scale prediction fusion method is employed to utilize global information at various scales, and the ECA attention module is employed to aggregate attention information from each layer in the compression path. Extensive experiments demonstrate that the proposed method achieves performance comparable to or better than networks such as Deeplabv3+, FCN, and UNet++.