A Residual Optronic Convolutional Neural Network for SAR Target Recognition

Ziyu Gu, Zicheng Huang, Xiaotian Lu, Hongjie Zhang, Hui KuangPhotonics 2025 · DOI

TL;DR

This paper introduces a novel optronic convolutional neural network (res-OPCNN) that leverages light to perform almost all computations for Synthetic Aperture Radar (SAR) target recognition. This approach significantly reduces computational costs and power consumption compared to traditional digital deep learning methods, while achieving comparable accuracy.

What Problem Does This Paper Address?

Deep learning (DL) has revolutionized fields like remote sensing and automatic target recognition (ATR), particularly for Synthetic Aperture Radar (SAR) imagery. However, the immense computational demands and power consumption of these digital DL models pose significant challenges, hindering real-time processing and widespread deployment, especially in resource-constrained environments. This is exacerbated by the fact that the growth in computational needs for DL models outpaces the improvements predicted by Moore's Law for electronic processors. The paper addresses this by exploring optical neural networks (ONNs) as a more efficient alternative.

Methods

The core innovation is the development of a residual optronic convolutional neural network (res-OPCNN). This architecture combines optical components for computation with electronic control for training and parameter adjustment.

Key methodological aspects include:

  • Optronic Convolutional Layers: The network utilizes optical systems (specifically, 4f lens systems and spatial light modulators - SLMs) to perform the convolution operations, which are computationally intensive in digital networks. This allows for processing at the speed of light.
  • Residual Connections: To combat the vanishing/exploding gradient problem common in deep networks, the authors incorporate optical residual connections. These connections allow the signal to bypass certain convolutional layers, providing a more direct path for gradient flow and improving training stability.
  • Optical Down-Sampling: To reduce the spatial dimensions of feature maps (a common practice in CNNs for efficiency), an optical demagnification system using lenses is employed.
  • Global Average Pooling (GAP) in Optics: The classification stage uses an optical GAP layer to aggregate feature information, transforming it into a confidence vector. The intensity of the Airy disk produced by focusing the light is used to represent this confidence.
  • Encoder-Classifier Structure: The network is divided into an Encoder (for feature extraction using optical convolutions and down-sampling) and a Classifier (for mapping features to target classes using optical GAP).

Datasets

The Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset was used for training and validation. This dataset is a standard benchmark for SAR target recognition tasks. The paper utilized 2747 images for training and 2425 images for validation.

Key Insights and Ideas

  • Optics for Computation is Viable for ATR: The paper demonstrates that optical methods can effectively perform complex operations like convolution and down-sampling, crucial for SAR target recognition, with significant efficiency gains.
  • Residual Connections Enhance Optical Networks: Adapting the successful residual learning concept from digital DL to an optronic architecture proved effective in improving training stability and performance.
  • Significant Power and Computational Savings: The primary "aha moment" is the dramatic reduction in computational complexity and power consumption achieved by shifting computations to optics. This opens doors for real-time, low-power ATR systems.
  • Lightweight and Feasible: The res-OPCNN is presented as a "lightweight" and feasible approach, suggesting that optronic solutions are not just theoretical but practical for real-world applications.

Key Results

  • Accuracy: The res-OPCNN achieved a classification accuracy of 95.3% on the MSTAR dataset, comparable to digital CNNs.
  • Computational Cost: The optronic approach required only 4% of the computational costs of a comparable digital CNN (748.6 KMac vs. 18.56 MMac).
  • Power Consumption: Power consumption was reduced by over 75%, with the optronic system consuming at most 250 W compared to 1000 W for the digital CNN.

Limitations and Future Work

The authors acknowledge several limitations and suggest future research directions:

  • Speckle Noise: The MSTAR dataset, like many SAR datasets, suffers from speckle noise. Future work should focus on developing measures to suppress this noise for improved performance.
  • Down-Sampling Robustness: The current optical down-sampling method is sensitive to positional variations. Exploring more robust down-sampling techniques that are less susceptible to misalignment is crucial for portability.
  • Miniaturization and Integration: For practical applications, further miniaturization and integration of the optronic components are necessary. This is highlighted as a key development trend for optical neural networks (ONNs).

Summary updated

Abstract

Deep learning (DL) has shown great capability in remote sensing and automatic target recognition (ATR). However, huge computational costs and power consumption are challenging the development of current DL methods. Optical neural networks have recently been proposed to provide a new mode to replace artificial neural networks. Here, we develop a residual optronic convolutional neural network (res-OPCNN) for synthetic aperture radar (SAR) recognition tasks. We implement almost all computational operations in optics and significantly decrease the network computational costs. Compared with digital DL methods, res-OPCNN offers ultra-fast speed, low computation complexity, and low power consumption. Experiments on the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset demonstrate the lightweight nature of the optronic method and its feasibility for SAR target recognition.
BibTeX
@inproceedings{gu2025residual,
  author = {Ziyu Gu and Zicheng Huang and Xiaotian Lu and Hongjie Zhang and Hui Kuang},
  title = {{A Residual Optronic Convolutional Neural Network for SAR Target Recognition}},
  booktitle = {Photonics},
  year = {2025},
  doi = {10.3390/photonics12070678},
}
Figure