Light Weight Solution for Stem and Leaf Classification in Tea Industry, Hybrid Color Space for Black Tea Classification

Abstract

This research proposes a new approach for stem and leaf classification in tea industry by deriving new color components which is simple in implementation, high in accuracy and low in cost than the multilayer neural network approaches. It has been used 270 set of tea stem and leaf sample in order to get 95% accuracy and the images were captured using a DSLR Nikon D3100 camera under controlled light condition. This paper includes an algorithm to pre-process images using image processing algorithms such as Otsu algorithm for threshold detection and Moore-Neighbor tracing algorithm for contour detection. Furthermore, it has been proposed a solution to select color components from existing color spaces which have highest discriminating power, deriving new color components by applying feature selection algorithms and calculating classification threshold and accuracy for each feature. The threshold values of the classification points will be used to differentiate stems and leaves as a single layer neural network, which is more lightweight than multi-layer neural network, which will also give a higher accuracy.