Comparative Analysis of CNN and ResNet in Automatic Modulation Recognition

Automatic modulation recognition (AMR) is a crucial task in wireless communication systems, enabling intelligent signal processing and spectrum management. Among various deep learning methods, Convolutional Neural Networks(CNNs) and Residual Networks(ResNets) consistently outperform traditional approaches in AMC tasks. However, the effectiveness of residual structure in shallow networks is still worth studying. With both networks constrained to four layers, this study conducts a comprehensive experimental study comparing CNN and ResNet models for AMC using the RML2018 dataset, a widely adopted benchmark for modulation classification. Both networks are trained under identical settings (epoch=200, batch_size=128) to ensure a fair comparison. The results reveal that ResNet's classification accuracy degrades to random-guessing levels at -10dB SNR, performing significantly worse than CNN. As the Signal to Interference plus Noise Ratio(SNR) increases to 20dB, the gap between the two gradually narrows, though ResNet still slightly trails behind CNN. However, ResNet demonstrates superiority in high SNR scenarios. This finding contradicts the prevailing assumption that ResNet universally outperforms other architectures in deep learning applications. This paper analyzes potential reasons for this phenomenon, including residual connection design and feature extraction efficiency.