An Efficient Deep Learning Model for Automatic Modulation Recognition

Abstract

Automatic Modulation Classification (AMC) has emerged as a critical research domain with wide-ranging applications in both civilian and military contexts. With the advent of artificial intelligence, deep learning techniques have gained prominence in AMC due to their unparalleled ability to automatically extract relevant features. However, most contemporary AMC models rely heavily on downsampling strategies to increase the receptive field while reducing computational complexity. Empirical evidence indicates that progressive downsampling substantially reduces the spatial resolution of feature maps, leading to poor generalization, particularly for closely related modulation schemes. To address these challenges, this paper proposes a novel Multiscale Dilated Pyramid Module (MDPM). In contrast to traditional downsampling techniques, MDPM mitigates resolution loss and retains a broader range of features, facilitating more comprehensive recognition. Furthermore, the multiscale features captured by MDPM enhance the robustness of the model to noise, thereby improving classification performance in noisy environments. The model's efficiency is further optimized through the integration of group convolutions and channel shuffle techniques. Extensive experimental results and evaluations confirm that the MDPM-based approach surpasses state-of-the-art methods, underscoring its significant potential for practical deployment. The signal data¬base and model can be freely accessed at https://pan.baidu.com/s/1g_HQXcRXshrT8nwKUNDYrQ?pwd=9ug6.