YOLO Network-Based Extraction of Target Geometry From Time-Frequency Representation

Abstract

In response to challenges in target geometric parameter extraction from radar echoes, an improved YOLOv8 neural network model is proposed in this paper to recognize target types and extract geometric parameters of targets from time-frequency representation (TFR). Compared with the existing network model, the ability to extract geometric parameters is enhanced by improving the network structure, integrating the attention mechanism, and introducing deformable convolution operators. The attributed scattering center (ASC) models for targets are established to generate the TFRs. This effectively addresses the challenge of generating target datasets with varying geometric parameters for the neural network. The robustness of the network is verified through comparison experiments and ablation experiments for different types of targets. Experimental results, including metric functions and data on reconstructed geometric structures, demonstrate good consistency between the extracted and actual geometric parameters, and the trained network model has a good generalization ability for more complicated structural targets. This study explores the possibility of applying neural networks with TFR to extract target parameters.

Keywords

• Neural network
• scattering center (SC)
• time-frequency representation (TFR)
• parameter extraction