Background Pixel Removal for Near-Infrared Hyperspectral Images Based on the Pixel-Wise Standard Deviation of Reflectance.

Abstract

This study proposes a method to remove background pixels from near-infrared hyperspectral images based on the pixel-wise standard deviation of reflectance method (px-wise SD method). This method calculates the standard deviation (SD) of reflectance in each pixel, namely each spectrum, and determines a threshold to distinguish between background and object pixels from the resulting histogram of the px-wise SD. The method effectiveness is evaluated using hyperspectral images of a leaf-like pastry with a hole placed on either a low-reflectance sheet or white paper. On white paper, the px-wise SD of reflectance exhibits a trimodal histogram with two prominent peaks and one small peak between them. The prominent peak with a lower SD corresponds to the white paper pixels, whereas the other peak with a higher SD is associated with the surface and edge pixels of the pastry. The small peak represents the pixels of the hole. The background and object pixels can be effectively separated by setting a threshold between this small peak and the prominent peak for the pastry pixels. Moreover, the mean spectrum calculated using only object pixels remains consistent, regardless of the type of background material. Conversely, the mean spectrum calculated using all pixels is distorted due to the spectral inclusion of the background material.