Wheat Head Recognition Based On Object Detection And Ensemble Learning

Abstract

In the context of the Vietnamese government wanting to develop a domestic wheat agriculture to overcome the fact that imported wheat is currently high in price and low in availability, in this thesis, we propose an application that aids in wheat head recognition. The application is created based on object detection and ensemble learning. Firstly, two individual YOLOv5 models of different weights are trained and used to project bounding boxes of objects that are possible wheat heads in an image or video. Secondly, using Weighted Box Fusion algorithm as the post-process, the resulting boxes generated in the previous step are fused together and the newly fused boxes’ confidence scores are recalculated based on the scores of the original boxes that made them up. In short, this thesis aims to utilize the YOLOv5 object detection models and the Weighted Box Fusion approach to recognize wheat head by wheat head, as well as identify the crop distribution across an area.