Semantic Segmentation Using U-Net Deep Learning Network for Quince Phenotyping on RGB and HyperSpectral Images

Abstract

Semantic segmentation based on the deep learning techniques can be used for the non-invasive phenotyping of quinces. In this paper we present a deep neural network for generating pixel wise mask from RGB and Hyperspectral images of the quinces using the U-Net architecture. The generated mask will be very useful for the experts involved in the phenotyping in order to get the dimension of the quinces. Also it can be used in the future for automatic plucking of quinces by the robot. This paper also compares the evaluation metrics of the model trained on both RGB and HSI data. We were able to achieve an accuracy of 93.33% and 70.225% for HSI and RGB data respectively. The developed segmentator is freely available in the GIT repository. The future works will include the model for detecting the ripeness of the quinces from the HSI data and also HSI images will be included in the dataset which will be helpful for the experts who are making research for other fruits.