A Modulation Recognition Algorithm Based on Multi-Scale Feature Enhancement and Attention Mechanism

Automatic modulation recognition （AMR） plays an important role in noncooperative communication systems. In recent years， deep learning-based automatic modulation recognition （DL-AMR） methods have achieved notable improvements in recognition accuracy and robustness compared to traditional approaches. However， existing DL-AMR methods generally suffer from limited feature extraction capabilities and difficulty in balancing accuracy and the complexity of the computational model. To address these limitations， we propose a modulation recognition architecture referred to as a multi-scale feature enhancement and attention mechanism network （MFEA-Net）. The proposed network employs convolution kernels with different receptive fields to extract multiscale signal features， which are then further compressed and fused through convolutional layers. For feature enhancement， depthwise convolution （DW-Conv） and dilated convolution （D-Conv） were used to replace standard convolutions to strengthen multiscale representations while reducing the number of parameters and computational cost. Squeeze-and-excitation （SE） attention was adopted to emphasize the informative channel features. Moreover， multi-head self-attention （MHSA） was introduced to model global dependencies at the sequence level. Global maximum pooling （GMP） and global average pooling （GAP） were combined to capture the global statistical features， and the final recognition results were produced through fully connected layers. Comparative experimental results show that， in the SNR≥0 dB range on both the RML2016.10a and RML2016.10b datasets， MFEA-Net achieves recognition accuracies exceeding 90% for the vast majority of modulation types. In addition， the overall accuracy （OA） of MFEA-Net respectively reached 63.01% and 64.61% on the two datasets despite the relatively low complexity of the model. Furthermore， the results of ablation studies verified the contributions of each module to the overall performance and thus demonstrate the rationality of the proposed architecture. Thus， the results demonstrate that MFEA-Net achieved a good balance between performance and complexity and showed promise for applications in resource-constrained scenarios.