Sorting different types of rice using smartphone based multispectral images

Spectral imaging techniques are important tools in chemical analysis. These tools combine the spectroscopic attributes of chemical measurements with the ones of imaging. Challenging applications of spectral imaging can be found in chemistry, biology, medicine, food science or agriculture, at both the micro and macro scales. In spectral imaging techniques, a spectrum is measured per pixel and the sample can be scanned without any kind of preparation [1]. Hyperspectral images (HSIs) are often composed of large number of pixels because one usually aim to cover large scenes or/and to get sufficient spatial resolution. In addition, a large number of variables are usually measured, to get sufficient spectral resolution and chemical selectivity. As a consequence, large data sets are generated, their analysis is time consuming, and high-performance computers might be required [2]. In addition, measuring too many collinear variables might translate into error propagation and over-fitting issues in multivariate modeling. A multispectral image is one that captures image data within specific wavelength ranges. The wavelengths can be separated by filters or detected via the use of instruments that are sensitive to particular wavelengths. The size of the multispectral images can be very small but would carry more information than simple RGB image. Rice, as one of the popular nutritional products, is of high amount of thought to be classified by nondestructive methods [3]. In this study we utilized a simple set-up including a laptop’s monitor and a smartphone to collect multispectral data so that the color of the laptop screen changed (9 colors) each time the object was photographed. Each image is taken in a separate spectral band in the visible spectral region. Then, multi-spectral images recorded from different Iranian rice were used in rice sorting using Partial Least-Squares Discriminant Analysis (PLS-DA). Specificity, precision and sensitivity for each class revealed that the discrimination model is adequate.