Efficient Channel Attention (ECA) for Spatial attention

To capture spatial dependencies in the multivariable time series data, ECA enhances the representational capacity of deep CNN.

ECA module is efficient for spatial attention feature extraction and has been widely used in modern deep CNN architecture. By avoiding dimensionality reduction and employing a local cross-channel interaction strategy via one-dimensional convolution, ECA captures inter-channel dependencies with minimal computational overhead. Furthermore, it dynamically adjusts the convolution kernel size to adapt to varying network depths.

Due to its effectiveness, the ECA1DLayer is provided to integrate this layer in the MATLAB deep learning pipeline. The implementation of ECA1DLayer depends on the customization layer template available in define custom deep learning layer.