ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks

Short summary

ECA-Net introduces an Efficient Channel Attention module that achieves over 2% Top-1 accuracy gain on ResNet50 with minimal parameters (80 vs. 24.37M) and computation (4.7e-4 GFlops vs. 3.86 GFlops) by using local cross-channel interaction via 1D convolution without dimensionality reduction.

AI-generated from the title and abstract; the full text is not read.