Augmented design and analysis of computer experiments: a novel tolerance embedded global optimization approach applied to SWIR hyperspectral illumination design.

Janos C Keresztes,Jan Audenaert,Karlien D’Huys,Bart De Ketelaere,R John Koshel,Peter Goos,Wouter Saeys

doi:10.1364/oe.24.029380

Janos C Keresztes, Jan Audenaert + Show 5 more

Open Access

https://doi.org/10.1364/oe.24.029380

Copy DOI

Abstract

A novel meta-heuristic approach for minimizing nonlinear constrained problems is proposed, which offers tolerance information during the search for the global optimum. The method is based on the concept of design and analysis of computer experiments combined with a novel two phase design augmentation (DACEDA), which models the entire merit space using a Gaussian process, with iteratively increased resolution around the optimum. The algorithm is introduced through a series of cases studies with increasing complexity for optimizing uniformity of a short-wave infrared (SWIR) hyperspectral imaging (HSI) illumination system (IS). The method is first demonstrated for a two-dimensional problem consisting of the positioning of analytical isotropic point sources. The method is further applied to two-dimensional (2D) and five-dimensional (5D) SWIR HSI IS versions using close- and far-field measured source models applied within the non-sequential ray-tracing software FRED, including inherent stochastic noise. The proposed method is compared to other heuristic approaches such as simplex and simulated annealing (SA). It is shown that DACEDA converges towards a minimum with 1 % improvement compared to simplex and SA, and more importantly requiring only half the number of simulations. Finally, a concurrent tolerance analysis is done within DACEDA for to the five-dimensional case such that further simulations are not required.

Highlights

In the past decade, hyperspectral imaging (HSI) has been demonstrated to be a more effective food quality inspection technology than traditional machine vision techniques, especially when defects cannot be observed in the visible (380–780 nm) spectral range [1,2,3,4,5]
In order to perform the simulations matching realistic orders of magnitude of the short-wave infrared (SWIR) case study, the goniometer measurements which were in photometric units needed to be scaled to SWIR radiometric units, which can be done with knowledge of the spectral power distribution
No angular-spectral dependencies were observed during the experiment,so we can assume that the integrated spectral radiance with fixed spectra is only dependent on the angular distribution

Summary

Introduction

Hyperspectral imaging (HSI) has been demonstrated to be a more effective food quality inspection technology than traditional machine vision techniques, especially when defects cannot be observed in the visible (380–780 nm) spectral range [1,2,3,4,5]. Short-wave infrared (SWIR) HSI has been shown to allow successful detection of fungi infestation in hazelnuts [5] and early detection of damage such as bruised apples [6,7], thanks to its potential for non-destructive chemical imaging. The prediction efficiency in these studies was limited by the illumination non-uniformity. The goal of this paper is to develop methods to simulate the performance of SWIR HSI illumination methods to maximize the performance of built systems. To date the characterization of the performance of such systems using meta heuristic optimization such as simplex or simulated annealing methods is time consuming, especially for non-convex problems [9]. During optimization, the proposed method provides simultaneously tolerance information of the entire design space

Objectives

Methods

Results

Conclusion