Semantic ghost imaging based on recurrent-neural-network.

Area of Science:

• Computational imaging
• Machine learning for optics

Background:

• Ghost imaging (GI) reconstructs images by correlating illumination patterns with detected intensities.
• Traditional GI methods suffer from slow imaging speeds and limited image quality due to independent sample processing.
• Existing techniques often struggle with low sampling rates, impacting practical applications.

Purpose of the Study:

• To introduce a novel recurrent neural network (RNN) system, termed GI-RNN, for ghost imaging.
• To explore the potential for non-linear connections between sequential samples in GI.
• To enhance image quality and reduce sampling rates in ghost imaging.

Main Methods:

• Developed a ghost imaging system based on a recurrent neural network (GI-RNN).
• Trained the GI-RNN model on datasets including MNIST handwriting and natural images.
• Evaluated image quality using metrics like dB improvement and Structural Similarity Index (SSIM).

Main Results:

• GI-RNN achieved significantly higher image quality compared to traditional basic correlation and compressed sensing algorithms at a 1.28% sampling rate.
• The system demonstrated a 12.58 dB improvement over basic correlation and 6.61 dB over compressed sensing for MNIST data.
• Trained GI-RNN showed strong generalization, recovering natural scenes with SSIMs > 0.7 at a 3% sampling rate.

Conclusions:

• The proposed GI-RNN effectively recovers high-quality images at remarkably low sampling rates.
• The RNN-based approach captures non-linear relationships in sequential GI data, overcoming limitations of linear methods.
• GI-RNN offers a promising solution for faster and higher-quality ghost imaging applications.