Text Region-Based Convolutional Neural Network for Precision Agriculture

Abstract

Application of Neural Networks in Precision Agriculture is now more widespread than ever. Neural networks have been extensively used in various tasks in precision agriculture, such as plant detection, disease detection, yield estimation, and soil classification. In this thesis, we build a blueberry plant image dataset for object detection with additional directional text that indicates the where in the image blueberry plant is. We train the Region-based Convolutional Neural Network (RCNN) model twice. First, using its original architecture that utilizes the Selective Search algorithm to create region proposals. Then, we modify the model by replacing Selective Search algorithm with additional text data to generate region proposals. Through performance analysis of both models on the test data, we show that the text method saves significant time on both training and inference while having good enough accuracy to compete with original model.