Efficient target recognition in ghost imaging via optimized 1D convolutional networks at a low sampling ratio

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Computer Vision

Background:

• Object recognition is crucial for imaging systems.
• Conventional imaging requires high data bandwidth.
• Computational ghost imaging (CGI) offers a low-bandwidth alternative but faces challenges in feature extraction.

Purpose of the Study:

• To demonstrate the efficacy of a convolutional neural network (CNN) for object recognition in both conventional and CGI.
• To evaluate the CNN's performance in an image-free CGI framework with subsampled data.
• To assess the CNN's resilience to noise and explore noise-augmented training strategies.

Main Methods:

• Training a CNN on the Fashion-MNIST dataset for image classification.
• Applying the trained CNN to CGI using compressed bucket measurements at an 80% sampling ratio.
• Investigating the CNN architecture's robustness against noise, including noise-augmented training.

Main Results:

• The CNN achieved 92.91% accuracy in conventional image-based classification.
• In an image-free CGI setting, the CNN attained 97.55% accuracy using only 80% subsampled data.
• Noise-augmented training significantly improved the CNN's performance in high-noise conditions.

Conclusions:

• CNNs are effective for object recognition in both conventional and CGI modalities.
• The proposed CNN architecture enables efficient, low-bandwidth optical sensing by extracting features from highly subsampled data.
• The CNN offers a flexible, noise-resilient solution for object recognition, bridging conventional imaging and computational optics.