Detection and classification in the compressed domain for multispectral images

Various applications would benefit from rapid inference on multispectral images at the point of sensing. However, the acquisition of a full-resolution multispectral image requires advanced spectrometers and prohibitive sensing time. Also, performing the high-level vision tasks such as classification and segmentation on the multispectral data consumes more computation power than on the common RGB images. 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. The further high-level analysis of images is performed on the reconstructed high-dimensional images. And a random sensing matrix may not be physically realizable or the best fit for extracting information pertaining to a high-level vision task. A realizable low-cost data acquisition scheme and a fast processing system that makes inference based on the acquired signal are desired for multispectral images. In this thesis, we present a systematic way to jointly optimize the sensing scheme subject to optical realizability constraints, and make inference of the multispectral image in the compressed domain. In the first part of the thesis, we state some open questions in compressed inference. We review the theory on inference in the compressed domain. We formulate the problem for compressed inference and state metrics to evaluate the inference performance. We then review some existing realizable optical compressed sensing imaging systems designed for multispectral images and derive the forward model of data acquisition. The feasibility of performing detection, classification and segmentation in the compressed domain directly is then discussed for the multispectral images. Using tools from detection and estimation theory, we derive the optimal decision rule to perform compressed detection, classification and segmentation in a simple data setting. Also, the feasibility of adjusting the optical acquisition schemes jointly with the neural network is discussed. The architecture of neural networks that can achieve the performance of the optimal decision rule is proposed and the existence of optimal weights is discussed. Next, we use a synthetic dataset to compare the performance of the proposed neural network and the optimal decision rule. Several synthetic multispectral image datasets and a clinical tumor biopsy dataset are used 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.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2021

Author dc:creator

• Li, Yuqi

Contributors dc:contributor

• Bresler, Yoram

• compressed inference
• multispectral image
• deep neural network
• detection
• classification
• segmentation
• optimal decision rule

Statement dc:rights

• Copyright 2020 Yuqi Li

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/109401

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F109401