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

This paper addresses the trade-off between performance and complexity in channel attention for deep convolutional neural networks. Dissecting the channel attention module of SENet, the authors find empirically that avoiding dimensionality reduction is important for learning channel attention and that appropriate cross-channel interaction can preserve accuracy while cutting complexity. They propose the Efficient Channel Attention (ECA) module, which performs local cross-channel interaction without dimensionality reduction, implemented efficiently via 1D convolution with an adaptively selected kernel size.

ECA adds only a handful of parameters, for example about 80 additional parameters and 4.7e-4 GFLOPs against a ResNet50 backbone of 24.37M parameters and 3.86 GFLOPs, while boosting Top-1 accuracy by more than 2%. Extensive evaluation on image classification, object detection and instance segmentation with ResNet and MobileNetV2 backbones shows ECA is more efficient than competing attention modules while performing favorably against them.