Coupled deep learning coded aperture design for compressive image classification.

Area of Science:

• Computer Vision
• Machine Learning
• Optical Engineering

Background:

• Single-pixel cameras (SPCs) offer a cost-effective alternative to traditional imaging systems.
• Compressive sensing architectures in SPCs often require complex reconstruction algorithms.
• Accurate classification of single-pixel measurements remains a challenge.

Purpose of the Study:

• To develop a coupled deep learning framework for simultaneous coded aperture design and classification in SPCs.
• To eliminate the image reconstruction step in SPC-based classification tasks.
• To improve classification accuracy and computational efficiency for compressive measurements.

Main Methods:

• A unified deep neural network was trained to co-optimize the binary sensing matrix of an SPC and a classification network.
• Two distinct network architectures were proposed: one for re-projecting measurements to image size, and another for direct feature extraction from compressive measurements.
• The approach was validated through simulations on two image datasets and a physical test-bed implementation.

Main Results:

• The proposed coupled deep learning method successfully performs classification directly from single-pixel measurements, bypassing reconstruction.
• One proposed network architecture achieved approximately 10% higher classification accuracy compared to state-of-the-art methods.
• The second network architecture demonstrated a 2x increase in computational speed.

Conclusions:

• The coupled deep learning approach offers a reliable and efficient alternative for classification using SPCs.
• Simultaneous optimization of sensing and classification within a neural network framework is a promising direction for compressive imaging.
• The developed methods show potential for real-world applications requiring fast and accurate classification from limited measurements.