Towards constructing a DOE-based practical optical neural system for ship recognition in remote sensing images

Abstract

The optical neural network system based on the 4f system (4f-ONN) is a feasible solution for on-orbit real-time target detection and recognition on remote sensing images, as it can directly modulate and encode the two-dimensional image information. However, traditional 4f systems based on spatial light modulators (SLMs) often encounter misalignment errors during assembly due to the reflective optical path used in SLMs. Additionally, implementing electronic SLM in space-based applications introduces problems such as particle number reversal, significantly reducing the reliability of the system. To address these issues, this paper proposes the adoption of diffractive optical elements (DOEs) to construct the 4f-ONN system. The DOE-based design offers a more compact structure, enhanced reliability, and reduced energy consumption, making it highly suitable for on-orbit image processing applications. Due to the expensive and time-consuming nature of the DOE manufacturing process, a design approach for a 4f system based on DOE was pursued through software simulation in this paper. The simulation phase involved the utilization of electronic neural networks to acquire the physical parameters of the DOE mask, while incorporating the array theorem and Fraunhofer diffraction theorem to accurately calculate the physical dimensions of the DOE. The effectiveness of the adopted DOE-based 4f system was initially validated through experiments on a simple pattern dataset. Subsequently, simulation experiments were conducted on three public datasets, namely Mnist, Fashion-mnist, and QuickDraw16, to confirm the efficacy of the DOE-based 4f system in classification tasks. Lastly, target recognition experiments were performed on the GF-2 dataset, and a corresponding hardware system was developed to demonstrate the potential of the DOE-based 4f system in on-orbit image processing.