Adaptive Refocusing and Target Recognition in SAR Images Using an Early Exit Strategy

Synthetic Aperture Radar (SAR) is easily affected by phase errors, resulting in blurred and distorted images, which increases the difficulty of target recognition tasks. Traditional SAR autofocus methods typically rely on iterative optimization strategies for image restoration, which demand significant computational resources. This paper presents an Adaptive Convolutional Neural Network (A-CNN), integrating an adaptive autofocus model with an Early Exit (EE) mechanism and a complex feature extraction network. This approach not only attains high recognition accuracy but also exhibits rapid convergence. The adaptive autofocus model with EE can update its model parameters adaptively, achieving precise compensation for SAR images with different levels of phase errors. For SAR data in complex format, the complex feature extraction network is designed to fully capture phase information, facilitating the analysis and interpretation of temporal and structural features within SAR images. Experimental results demonstrate that ACNN outperforms other network structures in terms of recognition accuracy and convergence speed.