EmbracingLimitedImperfectDatasets

Embracing limited and imperfect datasets in plant disease recognition using deep learning.

Currently maintained by Mingle Xu.

We are collecting public plant disease datasets in PPDRD project.

Motivation

Recent advancements in deep learning have brought significant improvements to plant disease recognition. However, achieving satisfactory performance often requires high-quality training datasets, which are challenging and expensive to collect. Consequently, the practical application of current deep learning-based methods in real-world scenarios is hindered by the scarcity of high-quality datasets. Consequently, we argue that embracing non-high-quality datasets is more convincing and practical. Although this embrace brings opportunities, new challenges exist. A taxonomy of related challenges is, therefore, proposed to enrich our understandings. With this perspective, we do hope that deep learning can be implemented in real-world applications with satisfactory performance.

Taxonomy of the challenges

• Limited data: the training dataset is not in large-scale.
• • Class-level: consider the variation of different classes within the training dataset.
• • • : All classes have similar few annotated images, where trained models may get low performance for all classes.
• • • : One class has many more annotated images than another class, where trained models may get high performance in the former class but suffer in the latter class.
• • Dataset-level: consider the diversity between the training and test datasets.
• • • : The training and test datasets share the same label spaces but are in different distribution spaces, where trained models may get low test performance.
• • • : Unknown (new) classes exist in the test dataset, where trained models will consider the corresponding image into a known class and not distinguish the unknown from known classes.
• Imperfect data: the training dataset is annotated in an undesired way.
• • : Training datasets have labeled and unlabeled images simultaneously, where utilizing the unlabeled images may contribute to the test performance.
• • : Training datasets are given with only coarse-grained annotations, where utilizing these annotations is challenging to train models.
• • : Some annotations may be inaccurate, where it is challenging to get decent test performance by utilizing these annotations to train models.

Related research