Style-Consistent Image Translation: A Novel Data Augmentation Paradigm to Improve Plant Disease Recognition

Area of Science:

• Computer Vision
• Plant Pathology
• Machine Learning

Background:

• Deep learning excels in plant disease recognition but requires substantial annotated data, which is often scarce or imbalanced in real-world scenarios.
• Existing data augmentation techniques struggle to create desirable variations for limited data and are not well-suited for object detection and instance segmentation tasks.
• Current methods typically generate variations within a single class, limiting their potential to improve model generalization.

Purpose of the Study:

• To develop a novel data augmentation paradigm that facilitates cross-class variation transfer for scarce and imbalanced plant disease datasets.
• To design a data augmentation method that supports image classification, object detection, and instance segmentation tasks simultaneously.
• To address the limitations of current data augmentation in generating diverse variations and accommodating detection/segmentation tasks.

Main Methods:

• Proposed a novel data augmentation paradigm that transfers variations from a source class to a target class, preserving domain-invariant features.
• Leveraged prior masks as input to enable the reuse of original annotations, facilitating application to object detection and instance segmentation.
• Collected a dataset of 1,258 tomato leaf images with 1,429 instance segmentation annotations, covering five diseases and healthy samples.

Main Results:

• The proposed cross-class data augmentation effectively generates desirable variations, improving model performance.
• The method successfully supports image classification, object detection, and instance segmentation tasks simultaneously.
• Experimental results demonstrate significant performance gains for diverse deep learning-based methods on the collected tomato leaf dataset.

Conclusions:

• The novel data augmentation paradigm offers a powerful solution for addressing data scarcity and imbalance in plant disease recognition.
• The approach enhances the applicability of data augmentation to object detection and instance segmentation, expanding its utility.
• This work contributes to advancing deep learning applications in agriculture for more accurate and efficient crop monitoring.