Scholarlink

<p>Determining an appropriate Coherent Processing Interval (CPI) is a fundamental challenge in Inverse Synthetic Aperture Radar (ISAR) imaging, as it directly governs the trade-off between cross-range resolution and rotational motion–induced image blurring. Since the CPI depends on both the Pulse Repetition Interval (PRI) and the number of transmitted pulses, this paper focuses on the adaptive determination of the optimal pulse number under a fixed-PRI assumption. A Double Deep Q-Network (Double DQN) based reinforcement learning framework is proposed, in which an agent sequentially adjusts the number of pulses used for image formation in order to maximize ISAR image quality. At each time step, the agent observes the formed ISAR image, extracts discriminative features using a convolutional neural network, and selects a discrete action to increase, decrease, or maintain the current pulse number. Image quality is quantitatively defined using a sparsity-based metric, which reflects image focus and concentration of dominant scatterers, and a Gaussian-shaped reward function is designed to guide the learning process. Through this formulation, the agent implicitly balances two CPI-dependent physical effects: insufficient target rotation leading to poor cross-range resolution and excessive rotation leading to image blurring. The proposed method is evaluated using simulated ISAR data for point-scatterers-based rigid body targets with varying rotational velocities and initial pulse numbers conditions. Experimental results indicate that the trained agent reliably converges to an optimal or near-optimal number of pulses, yielding well-focused ISAR images without the need for explicit cross-range resolution enhancement or rotational motion compensation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Novel Deep Reinforcement Learning-Based Approach for Optimal Pulse Number Selection in ISAR Imaging

  • 期刊论文

  • 发表日期：

    2026年6月30日

查看全文
  • Reza Mohammadi,
  • Ali Jabar Rashidi

摘要

Determining an appropriate Coherent Processing Interval (CPI) is a fundamental challenge in Inverse Synthetic Aperture Radar (ISAR) imaging, as it directly governs the trade-off between cross-range resolution and rotational motion–induced image blurring. Since the CPI depends on both the Pulse Repetition Interval (PRI) and the number of transmitted pulses, this paper focuses on the adaptive determination of the optimal pulse number under a fixed-PRI assumption. A Double Deep Q-Network (Double DQN) based reinforcement learning framework is proposed, in which an agent sequentially adjusts the number of pulses used for image formation in order to maximize ISAR image quality. At each time step, the agent observes the formed ISAR image, extracts discriminative features using a convolutional neural network, and selects a discrete action to increase, decrease, or maintain the current pulse number. Image quality is quantitatively defined using a sparsity-based metric, which reflects image focus and concentration of dominant scatterers, and a Gaussian-shaped reward function is designed to guide the learning process. Through this formulation, the agent implicitly balances two CPI-dependent physical effects: insufficient target rotation leading to poor cross-range resolution and excessive rotation leading to image blurring. The proposed method is evaluated using simulated ISAR data for point-scatterers-based rigid body targets with varying rotational velocities and initial pulse numbers conditions. Experimental results indicate that the trained agent reliably converges to an optimal or near-optimal number of pulses, yielding well-focused ISAR images without the need for explicit cross-range resolution enhancement or rotational motion compensation.