Automatic Modulation Recognition: An Efficient Feature-Mixing Capsule Replaces the Fully Connected Layer

In recent years, numerous deep learning (DL) approaches have been applied in automatic modulation recognition (AMR), achieving remarkable success. However, existing models primarily rely on fully connected layers for the detection of the final result, leading to significant redundancy and failing to fully utilize the positional correlation of extracted high-dimensional features. This limitation is particularly critical for low signal-to-noise ratio (SNR) signals, where features are not as distinct. This paper introduces an efficient feature-mixing homogeneous capsules (EFMixer-HCs) structure to replace fully connected layers. Capsules are created along the feature direction, with each capsule representing a vector whose values capture local feature attributes. These capsules are then fused to fully utilize positional correlations across features, enabling the model to learn complex patterns and fine-grained details more efficiently. Finally, extensive evaluations are conducted on datasets covering diverse application scenarios. Compared to fully connected structures, models incorporating the proposed structure achieve an average accuracy improvement of 0.19%–16.4%, while the number of model parameters is reduced by up to 56%.