Lightweight Deep Neural Networks for Ship Target Detection in SAR Imagery

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Deep convolutional neural networks (DCNNs) are effective for synthetic aperture radar (SAR) ship detection.
• High computational and storage demands of DCNNs hinder their use on resource-constrained platforms like satellites.

Purpose of the Study:

• To propose lightweight detection networks for SAR ship target detection.
• To address the limitations of DCNNs in terms of storage and computational cost for onboard applications.

Main Methods:

• Designed a network structure optimization algorithm based on the multi-objective firefly algorithm (NOFA).
• NOFA encodes filters of a high-performing network into probabilities to guide the inheritance of filter structures and parameters.
• Employed multi-objective firefly optimization (MFA) to optimize the probability list for balancing detection precision and network size.
• Utilized network pruning technology to convert optimized encodings into lightweight detection networks.

Main Results:

• The proposed method generates a set of lightweight networks tailored to different precision-size trade-offs.
• Experiments on SSDD and SDCD datasets demonstrate the effectiveness of the developed lightweight networks.
• The networks offer greater flexibility and reduced size compared to traditional detection networks.

Conclusions:

• The NOFA-based approach successfully creates efficient and lightweight DCNNs for SAR ship detection.
• This method enables the deployment of advanced ship detection capabilities on resource-limited airborne and spaceborne platforms.
• The optimized networks provide a practical solution for real-time SAR image analysis.