Abstract

We propose the use of multiplexed illumination to enable the classification of visually similar objects that cannot normally be distinguished. We construct a compact red, green, blue, and near-infrared light stage and develop a method to jointly select informative illumination patterns and training a classifier that uses the resulting images. We use the light stage to model training samples and synthesize noise-accurate images that drive the training process and a time-efficient greedy pattern selection scheme. The system delivers fast, accurate classification of previously indistinguishable samples, outperforming fixed-illuminant and conventional noise-optimal patterns. This work has potential applications spanning forgery detection and quality control in agriculture and manufacturing.

© 2021 Optical Society of America

Algorithm for training the minimum error one-class classifier of images

J. T. Guillen-Bonilla, E. Kurmyshev, and E. González
Appl. Opt. 47(4) 541-547 (2008)

Single-pixel neural network object classification of sub-Nyquist ghost imaging

Jia-Ning Cao, Yu-Hui Zuo, Hua-Hua Wang, Wei-Dong Feng, Zhi-Xin Yang, Jian Ma, Hao-Ran Du, Lu Gao, and Ze Zhang
Appl. Opt. 60(29) 9180-9187 (2021)

Learned sensing: jointly optimized microscope hardware for accurate image classification

Alex Muthumbi, Amey Chaware, Kanghyun Kim, Kevin C. Zhou, Pavan Chandra Konda, Richard Chen, Benjamin Judkewitz, Andreas Erdmann, Barbara Kappes, and Roarke Horstmeyer
Biomed. Opt. Express 10(12) 6351-6369 (2019)

Neural-network classifiers for automatic real-world aerial image recognition

Shlomo Greenberg and Hugo Guterman
Appl. Opt. 35(23) 4598-4609 (1996)

Multi-objective optimization for structured illumination in dynamic x-ray tomosynthesis

Xu Ma, Hao Xu, Carlos M. Restrepo, and Gonzalo R. Arce
Appl. Opt. 60(21) 6177-6188 (2021)