Tea moisture content detection with multispectral and depth images

Abstract

In this research, multispectral and depth images were utilized for tea moisture content detection, the problems on leaf surface orientation and detection height were studied specifically. For the leaf surface orientation issue, multispectral images (25 bands) of the front surface and back surface of tea leaves were collected. Based on the spectra with same surface orientation, regression models of tea moisture content were established. The R2P values of LSSVR models reach 0.77 and 0.68 for the front surface and back surface, respectively. To distinguish the surface orientation of tea leaves, an LDA classifier was built based on spectral band ratio information. The overall classification accuracy reaches 87.8%. The distribution map of tea moisture content was successfully generated by importing the spectra into the classifier and the regression model. For the detection height issue, the multispectral image and depth image of tea leaves were collected simultaneously. First, an experiment was designed to figure out the attenuation coefficient of each band and the calibration model of detection height. Then, the detection height information was introduced into each pixel of the multispectral image by image registration. According to the detection height and calibration model, the spectrum of each pixel was calibrated. Finally, through importing the modified spectra into the classifier and regression model, the visual detection of tea moisture content was realized with detection height calibration. This research promoted the practicability of tea moisture content detection, and improved the visualization detection technology based on the fusion of multispectral image and depth image.