Spectral Convolution Feature-Based SPD Matrix Representation for Signal Detection Using a Deep Neural Network.

Area of Science:

• Signal Processing
• Machine Learning
• Computer Vision

Background:

• Convolutional Neural Networks (CNNs) excel in visual tasks due to hierarchical feature extraction.
• Symmetric Positive Definite (SPD) matrices offer robust statistical representations for sample discrimination.
• Existing deep learning methods for signal detection can be improved for low signal-to-clutter ratio (SCR) scenarios.

Purpose of the Study:

• To propose a novel deep neural network signal detection method utilizing spectral convolution features and SPD matrices.
• To transform signal detection into a binary classification problem using extracted local features.
• To evaluate the performance and superiority of the proposed method against existing techniques.

Main Methods:

• Local features are extracted from CNN models (two types applied).
• Extracted features are used to construct SPD matrices.
• A deep learning algorithm is applied to the SPD matrices for target signal detection.

Main Results:

• The proposed method successfully transforms signal detection into a binary classification task.
• Evaluations on simulated and semi-physical datasets demonstrate the method's effectiveness.
• Under low SCR conditions, the method achieves a 0.5-2 dB gain compared to traditional deep neural network spectral signal detection.

Conclusions:

• The proposed deep learning method based on spectral convolution features and SPD matrices is effective for signal detection.
• This approach offers significant performance improvements, particularly in challenging low SCR environments.
• The use of SPD matrices enhances the ability to distinguish target signals within complex data.