JR2net: a joint non-linear representation and recovery network for compressive spectral imaging

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Deep learning models are state-of-the-art for compressive spectral imaging (CSI) recovery.
• Current methods often use deep neural networks (DNNs) for non-linear mapping but may not optimize representation for the CSI problem.
• The detached training of convolutional autoencoder (CAE) networks can lead to suboptimal spectral image representations.

Purpose of the Study:

• To propose a novel network, JR2net, that jointly optimizes non-linear representation and spectral image recovery for CSI.
• To integrate representation learning and image recovery into a single, end-to-end trained optimization problem.

Main Methods:

• Developed a joint non-linear representation and recovery network (JR2net).
• JR2net employs an optimization-inspired network based on the alternating direction method of multipliers (ADMM) formulation.
• The network learns a low-dimensional representation and performs spectral image recovery simultaneously, trained end-to-end.

Main Results:

• Achieved improvements up to 2.57 dB in peak signal-to-noise ratio (PSNR) compared to existing methods.
• Demonstrated a significant speed enhancement, performing approximately 2000 times faster than state-of-the-art techniques.
• The joint optimization approach proved superior for spectral image recovery in CSI.

Conclusions:

• JR2net effectively links representation learning and recovery tasks for improved CSI performance.
• The proposed method offers a more optimal and efficient approach to spectral image recovery.
• End-to-end training of JR2net leads to superior results in terms of accuracy and speed.