Ladder: A software to label images, detect objects and deploy models recurrently for object detection

TL;DR

Ladder is a user-friendly software that streamlines image labeling, object detection model training, and deployment, demonstrated by its application in assessing wheat stripe rust severity in UAV images for precision agriculture.

Contribution

The paper introduces Ladder, a novel interactive recurrent framework software that simplifies object detection tasks from labeling to deployment with a GUI.

Findings

Achieved 72% accuracy for low severity rust detection

Achieved 50% accuracy for medium severity rust detection

Achieved 80% accuracy for high severity rust detection

Abstract

Object Detection (OD) is a computer vision technology that can locate and classify objects in images and videos, which has the potential to significantly improve efficiency in precision agriculture. To simplify OD application process, we developed Ladder - a software that provides users with a friendly graphic user interface (GUI) that allows for efficient labelling of training datasets, training OD models, and deploying the trained model. Ladder was designed with an interactive recurrent framework that leverages predictions from a pre-trained OD model as the initial image labeling. After adding human labels, the newly labeled images can be added into the training data to retrain the OD model. With the same GUI, users can also deploy well-trained OD models by loading the model weight file to detect new images. We used Ladder to develop a deep learning model to access wheat stripe rust in RGB (red, green, blue) images taken by an Unmanned Aerial Vehicle (UAV). Ladder employs OD to directly evaluate different severity levels of wheat stripe rust in field images, eliminating the need for photo stitching process for UAVs-based images. The accuracy for low, medium and high severity scores were 72%, 50% and 80%, respectively. This case demonstrates how Ladder empowers OD in precision agriculture and crop breeding.