TY - GEN
T1 - Optimizing Optical Compressed Sensing for Multispectral DNN-Based Image Segmentation
AU - Li, Yuqi
AU - Bresler, Yoram
N1 - Funding Information:
§This research was supported in part by ARO grant W911NF-15-1-0479. The authors are grateful to Dr. Luke Pfister and Mr. Kianoush Falahkheirkhah for providing the initial idea of the line selection scheme. We would like to thank Dr. Rohit Bhargava and Dr. Shachi Mittal for allowing us to use the tumor biopsy dataset. The author’s interest in direct inference in the compressed domain was inspired by work with Dr. Eric Shields of Sandia National Laboratories.
Publisher Copyright:
© 2020 IEEE.
PY - 2020/11/1
Y1 - 2020/11/1
N2 - In various medical applications, the acquisition of a full-resolution multispectral image requires advanced spectrometers and prohibitive sensing time. Instead, compressed sensing (CS) circumvents this sensing process usually using a random sensing matrix to acquire fewer measurements and reconstructs the multispectral image based on a sparsity assumption. However, a random matrix may not be physically realizable nor the best fit for extracting information pertaining a high-level vision task such as segmentation. Here, we use a deep neural network to jointly optimize the sensing scheme subject to optical realizability constraints, and segment the multispectral image in the compressed domain. We use a synthetic dataset and a tumor biopsy dataset to verify the improvement of the obtained sensing scheme and compare the performance of the neural network with that of a known optimal decision rule.
AB - In various medical applications, the acquisition of a full-resolution multispectral image requires advanced spectrometers and prohibitive sensing time. Instead, compressed sensing (CS) circumvents this sensing process usually using a random sensing matrix to acquire fewer measurements and reconstructs the multispectral image based on a sparsity assumption. However, a random matrix may not be physically realizable nor the best fit for extracting information pertaining a high-level vision task such as segmentation. Here, we use a deep neural network to jointly optimize the sensing scheme subject to optical realizability constraints, and segment the multispectral image in the compressed domain. We use a synthetic dataset and a tumor biopsy dataset to verify the improvement of the obtained sensing scheme and compare the performance of the neural network with that of a known optimal decision rule.
UR - http://www.scopus.com/inward/record.url?scp=85107758418&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85107758418&partnerID=8YFLogxK
U2 - 10.1109/IEEECONF51394.2020.9443293
DO - 10.1109/IEEECONF51394.2020.9443293
M3 - Conference contribution
AN - SCOPUS:85107758418
T3 - Conference Record - Asilomar Conference on Signals, Systems and Computers
SP - 636
EP - 640
BT - Conference Record of the 54th Asilomar Conference on Signals, Systems and Computers, ACSSC 2020
A2 - Matthews, Michael B.
PB - IEEE Computer Society
T2 - 54th Asilomar Conference on Signals, Systems and Computers, ACSSC 2020
Y2 - 1 November 2020 through 5 November 2020
ER -