# TDP-SAR: Task-Driven Pruning Method for Synthetic Aperture Radar Target Recognition Convolutional Neural Network Model

## TL;DR

This paper introduces TDP-SAR, a new method to simplify deep learning models for SAR target recognition while maintaining accuracy.

## Contribution

The novel TDP-SAR method uses frequency domain analysis for task-driven pruning in SAR target recognition models.

## Key findings

- TDP-SAR reduces model parameters by 17.7% when applied to a J-CNN model.
- The method maintains recognition reliability even with compressed models.
- TDP-SAR adapts to different SAR image qualities better than baseline models.

## Abstract

Synthetic aperture radar (SAR) target recognition plays a crucial role in SAR image interpretation. While deep learning has become the predominant approach for SAR target recognition, existing methods face practical deployment challenges due to excessive model complexity. In addition, SAR images are less understandable compared to optical images, which leads to greater difficulties in analyzing the target features of SAR images in the spatial domain. To address the above limitation, we propose a novel task-driven pruning (TDP-SAR) strategy. Unlike conventional pruning techniques that rely on generic parameter importance metrics, our approach implements frequency domain analysis of convolutional kernels across different processing stages of SAR target recognition models. In the experimental section, we use the MSTAR benchmark dataset to prove that the TDP-SAR can not only effectively compress the model size but also adapt to different quality SAR images compared to baseline architectures. Particularly when processing the joint convolutional neural network (J-CNN) model proposed in the previous study, the number of parameters decreased by 17.7% before and after pruning. This advancement facilitates the practical deployment of deep learning solutions in resource-constrained SAR interpretation systems while preserving recognition reliability.