An iterative noisy annotation correction model for robust plant disease detection.

Area of Science:

• Computer Vision
• Plant Pathology
• Machine Learning

Background:

• Object detectors for plant disease detection are sensitive to noisy training data and inaccurate annotations.
• High-quality annotated datasets are essential but costly and time-consuming to create.
• Real-world agricultural data often contains location noise in bounding box annotations.

Purpose of the Study:

• To develop a method for learning robust feature representations from plant disease images with inaccurate bounding boxes.
• To reduce the model's dependency on precise annotation quality.
• To mitigate the impact of noisy labels in agricultural datasets.

Main Methods:

• Analysis of real-world noisy annotation distributions.
• Implementation of a teacher-student learning paradigm to correct inaccurate bounding boxes.
• Teacher model rectifies noisy bounding boxes; student model learns robust features from corrected data.

Main Results:

• Achieved a 26% performance improvement on a noisy dataset when applied to the Faster-RCNN detector.
• Reached approximately 75% of fully supervised performance with only 1% of labels available.
• Demonstrated generalization to semi-supervised learning and auto-labeling.

Conclusions:

• The proposed teacher-student method effectively addresses real-world location noise in bounding box annotations for plant disease detection.
• This approach alleviates challenges posed by noisy data in precision agriculture, optimizing data labeling.
• Encourages lower-cost investigation into plant disease detection and intelligent agriculture.