Mangosteen Fruit Detection Using Improved Faster R-CNN

Abstract

Agriculture is progressively getting access to scientific and technological breakthroughs in the digital era. In order to achieve smart agriculture, huge farms must be monitored and managed using advanced technologies. Anticipating and grading ripe fruit at harvest may aid in lowering storage costs and capturing market demand. Monitoring the ripening phase of the fruit also helps in the management of input and output criteria, which has practical implications in the harvesting process such as calculating the proper amount of water and nutrients at the end of the harvest, reducing traditional labor force, resulting in financial and human resource savings. In this research, we present a technique for identifying and classifying the ripening stage of mangosteen in agricultural fields. The study employs a two-stage approach based on a Faster R-CNN, a deep neural network which uses Region Proposal Network to extract the image region containing the item for the classification and identification of the mangosteen’s location, and an improved RoI (Region of Interest) Pooling algorithm by adding a RoI Align layer to optimize feature data during training. We enhanced both of the speed and accuracy while processing huge and complicated data sets using the suggested methodology. When employing a dataset of 10,000 photos of ripe mangosteen, our model outperforms the one-stage approach in terms of accuracy while maintaining real-time speed.