Systematic approach for using hyperspectral imaging data to develop multispectral imagining systems: Detection of feces on apples

Abstract

The large size of data sets generated using hyperspectral imaging techniques significantly increases both the capability and difficulty of designing detection and classification systems. Of particular interest is the confluence with increasing use of multispectral imaging in machine vision, particularly in the area of food safety inspection. The purpose of this study was to develop a robust method for selecting one or two wavelengths for multispectral detection systems using hyperspectral data. The actual performance of detection algorithms in terms of true positives and false positives was used as optimization criteria. Detection of fecal contamination on apples is an important health safety issue. Prior observations suggest reflectance or fluorescence imaging in the visible to near-infrared can be used to detect such contamination. For this study, 1:2, 1:20, and 1:200 dilutions of dairy feces were applied to 100 Golden and 100 Red Delicious apples. Apples were imaged using a hyperspectral system, and a uniform power transformation was used to reduce inter-apple intensity variability. Detection was accomplished by applying a binary threshold to transformed single wavelength images and images construct using ratios or differences of images at two different wavelengths. Optimization criteria allowed for a maximum of three false positives. For reflectance imaging, maximum detection rates for 1:20 dilution spots on Golden and Red Delicious apples images were 100% and 62.5% using R816 − R697 and R784 − R738, respectively. For fluorescence imaging, maximum detection rates for 1:200 dilution spots on Golden and Red Delicious apples were 97.9% and 58.3% using F665/ F602 and F647/ F482, respectively. In all case, more concentrated dilution spots were detected at 100%. Maximum detection rates for Red Delicious apples required use of a Prewitt edge-detection filter. In comparison, tests of wavelengths and algorithms identified in previous studies using statistical methods such as principal component analysis produced lower detection rates, mainly due to problems with false positives. The procedures used for developing detection algorithms are not specific to detecting feces on apples, and it is theoretically easy to extend the results to detection schemes involving many wavelengths. The problem is the classical dilemma of rapidly increasing computational time. Still, given the costs of thoroughly testing a candidate detection algorithm, the time maybe warranted. Furthermore, as machine vision systems are often limited to one or two wavelengths due to practical considerations including cost, exhaustive search algorithms based-on optimizing the output of candidate detection algorithms should be cost-effective.