Spectral Classification using a Dual Optical Setup and Deep Neural Networks

Abstract

Spectral classification allows material labeling based on spectral information. Single-pixel cameras (SPCs) have been used as a low-cost solution for acquiring spectral images, providing high-resolution spectral and low-resolution spatial information. Also, diffractive optical cameras (DOCs) based on multilevel phase masks (MPMs) can acquire spectral features to perform classification tasks. Traditional spectral classification approaches have not incorporated SPCs and DOCs into a single optical architecture. This work proposes a dual optical system based on SPC and DOC for spectral classification. Specifically, the height map in the MPM and the deep neural network parameters are jointly learned from end-to-end (E2E) optimization. The proposed method contains an optical layer that describes the dual system, a fusion layer that estimates the spectral image, and a classification network that labels the materials over spectral datasets. The simulation results show an improvement of up to 3% in classification metrics compared to other optical architectures.

Keywords

• spectral classification
• single pixel camera
• diffractive optical camera
• multilevel phase maskt
• end-to-end optimization
• deep neural networks