Fractal Filter Networks for Sea State Classification Using Radar Sea Clutter Data

This paper introduces Fractal Filter Networks (FrFNet), a novel network architecture for signal classification that learns in the fractal domain, leveraging the integration of fractal signal processing and deep learning methodologies. The primary concept of FrFNet is substituting the initial multi-head attention module of the Transformer model with a fractal filter, thereby introducing fractal characteristics into the Transformer architecture. The FrFNet framework mainly comprises three components: fractal feature extractor, fractal feature mixer, and classifier module. To assess the effectiveness of the proposed model, experiments were conducted on the IPIX dataset for sea state classification. Results demonstrate that FrFNet achieves a classification accuracy of 97.3%, outperforming various Transformer-style and CNN-style models. Furthermore, FrFNet exhibits notable advantages in convergence speed over other models. This research offers novel insights and technical avenues for the integration of fractal theory and deep learning.