Machine-Learning Radar Super Resolution for Real-Time Angular Detail

The patent creates a trained neural network model that copies the computational logic of traditional super-resolution methods. During training, the model learns the transformation patterns from low-resolution to high-resolution radar images using traditional methods as ground truth. During inference, the trained model applies these learned patterns through feed-forward propagation, replicating the resolution enhancement effect without executing the computationally intensive traditional algorithms, thus achieving both high angular resolution and fast processing speed.

The patent replaces the mechanical/computational system of traditional iterative super-resolution algorithms with a machine learning-based neural network system. Instead of using complex mathematical optimizations and iterative computations, the system uses a trained neural network that has internalized the resolution enhancement transformations. This substitution transforms the problem from one of computational complexity to one of model inference, dramatically reducing processing requirements while maintaining or improving resolution enhancement effectiveness.

The trained neural network model copies the functional output of traditional super-resolution methods without copying their computational complexity. The model encapsulates the essential transformation logic during training and executes it efficiently during inference through simple feed-forward operations, achieving the same image quality improvement with minimal computing resource consumption compared to the ground truth generation process.

The patent performs the computationally intensive training process in advance, during which the neural network learns and stores the optimal transformation patterns for super-resolution. Once trained, the model can be deployed for real-time applications without requiring the same level of computational resources. The preliminary action of training consolidates the computational workload into a one-time process, enabling efficient subsequent inference operations.

The patent segments the radar data processing into distinct functional stages: first processing low-level radar data through a neural network to generate intermediate representations that preserve comprehensive information, then applying super-resolution transformation to enhance angular resolution. This segmentation allows each stage to optimize for its specific function - information preservation in the first stage and resolution enhancement in the second stage - thereby achieving both complete information capture and high angular resolution simultaneously.

The patent transforms the radar data from a low-resolution representation in the angular dimension to a high-resolution representation by applying super-resolution techniques. The neural network processes the input data and outputs an enhanced representation that adds fine-grained angular detail while preserving the comprehensive information from the low-level radar data. This dimensional transformation in the angular domain enables both information completeness and high resolution to coexist.