Submitted to: Sensing and Instrumentation for Food Quality and Safety
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 4/16/2008
Publication Date: 6/12/2008
Citation: Lee, K., Kang, S., Delwiche, S.R., Kim, M.S., Noh, S. 2008. Correlation analysis of hyperspectral imagery for multispectral wavelength selection for detection of defects on apples. Sensing and Instrumentation for Food Quality and Safety. 2(2):90-96.
Interpretive Summary: Currently, apple packing houses are heavily reliant on manual inspection for identification and removal of apples that are bruised, cut, decayed, and blemished. Although machine vision is also routinely used, this technology has generally been limited to sorting for size and color. With advancement of computer processing and spectral imaging technology, it is becoming possible to extend the use of vision-based automatic sorting to include the aforementioned surface conditions that affect the safety and quality of agricultural commodities, such as apples. This study has examined the use hyperspectral imaging to identify wavelength regions of the visible and near-infrared spectra that are particularly useful for detection of defects on apple surfaces. (Hyperspectral imaging combines the useful features of digital imaging and spectroscopy, such that a series of images of an inspected commodity are produced, each at a different wavelength of radiated light.) Although this form of imaging is used mostly in analytical research, the advantageous feature of hyperspectral imaging is that it identifies an optimal set of two-to-four wavelengths that can be used in a multispectral image system, which is directly amenable to online inspection and sorting. In this study, a set 70 Fuji apples with varying types and levels of surface defects were examined by hyperspectral imaging over a wavelength range spanning from blue (418 nm) to short near-infrared (918 nm) light. By means of correlating the spectral response to arbitrarily assigned values of representative portions of normal (value = 0) and defect (value = 1) regions on the apple surface, the best pairs of wavelengths were identified in two mathematical formulations: the ratio of images and their difference. When the best wavelength ratio pair (670 nm / 684 nm) was applied to the set of Fuji apples, 195 of the 211 defect regions were correctly identified. The best wavelength difference pair (828 nm – 684 nm) had a slightly higher misclassification rate. Beneficiaries of this research include the instrument manufacturers, who now have a simplified procedure for selecting wavelengths during the design of dedicated multispectral image systems. Eventually, with the availability of such systems, the apple packing industry will benefit by having an automatic, rapid and objective method for removal of damaged apples during sorting and packing operations.
Technical Abstract: Visible/near-infrared reflectance spectra extracted from hyperspectral images of apples were used to determine wavelength pairs that can be used to distinguish defect regions from normal regions on the apple surface. The optimal wavelengths were selected based on correlation analysis between the wavelength band ratio ('1 / '2) or difference ('1 - '2) and the assigned value for the surface condition (0 = normal, 1 = defect). Spectral images of whole apples at the selected wavelengths were used to validate the correlation analysis. The correlation coefficients obtained using the correlation analysis for band ratio and difference were 0.91 and 0.79, respectively. When applied to the set of apple images, the band ratio model correctly identified 195 of the 211 defects on a set of 70 Fuji apples containing at least one defect region. Thus, the correlation analysis was demonstrated to be a feasible method for selecting wavelength pairs for use in distinguishing defects from normal areas on apples.