Introduction
Automatic Modulation Recognition (AMR) is a foundational task in cognitive radio, enabling receivers to identify signal modulation types—such as BPSK, QPSK, or QAM—so they can be properly processed. However, traditional AMR methods often struggle in high-noise environments. Conventional constellation diagrams lose vital statistical information at low signal-to-noise ratios (SNR), appearing as indistinguishable clouds of points. This paper introduces the Image-based CNN with Attention Model (ICAM), a framework designed to overcome these limitations by leveraging density-based signal representation and attention mechanisms.
Methodology
The ICAM framework employs a novel data representation called Contour Stella Images. Instead of simple scatter plots, the system calculates the density of signal points using a sliding window and maps these densities to a specific color scale. This approach preserves the statistical distribution of the signal, making patterns discernible even under heavy noise.
The model architecture utilizes a ResNet18 backbone enhanced with a Convolutional Block Attention Module (CBAM). This "plug-and-play" module incorporates both channel and spatial attention. Channel attention identifies the most meaningful feature maps, while spatial attention directs the network to focus on signal clusters rather than the noisy background. To evaluate the model, a dataset of 11,000 images was generated across eight modulation categories (BPSK, 4ASK, QPSK, OQPSK, 8PSK, 16QAM, 32QAM, and 64QAM) with SNRs ranging from -6dB to 20dB.
Results and Discussion
The ICAM model demonstrated superior performance compared to standard AlexNet and ResNet implementations. At an SNR of 0dB, the model achieved an accuracy exceeding 99%. Crucially, at the challenging -6dB SNR level, ICAM outperformed the standard ResNet18 by approximately 3%, validating the effectiveness of the attention mechanism in filtering noise.
Interpretability analysis using Grad-CAM (Gradient-weighted Class Activation Mapping) confirmed that the model focuses on the core clusters of the signal. For instance, in BPSK signals, the attention mechanism successfully isolates the two primary clusters, effectively ignoring the surrounding noise cloud. This transformation of a statistical problem into a texture-recognition task allows the CNN to excel where traditional methods fail.
Conclusion
While ICAM significantly improves AMR performance, distinguishing between phase-based modulations (like QPSK and 8PSK) remains difficult at extremely low SNRs due to physical signal blurring. Future research will focus on more advanced architectures to further enhance recognition accuracy for similar spatial signatures and move toward real-time hardware implementation.
Summary updated