Ambiguity-aware semi-supervised learning for leaf disease classification.

Abstract

In deep learning, Semi-Supervised Learning is a highly effective technique to enhances neural network training by leveraging both labeled and unlabeled data. This process involves using a trained model to generate pseudo labels to the unlabeled samples, which are then incorporated to further train the original model, resulting in a new model. However, if these pseudo labels contain substantial errors, the resulting model's accuracy may drop, potentially falling below the performance of the initial model. To tackle the problem, we propose an Ambiguity-Aware Semi-Supervised Learning method for Leaf Disease Classification. Specifically, we present a per-disease ambiguity rejection algorithm that eliminates ambiguous results, thereby enhancing the precision of pseudo labels for the subsequent semi-supervised training step and improving the precision of the final classifier. The proposed method is evaluated on two public leaf disease datasets of coffee and banana across various data scenarios, including supervised and semi-supervised settings, with varying proportions of labeled data. The results indicate that our semi-supervised method reduces the reliance for fully labeled datasets while preserving high accuracy by utilizing the ambiguity rejection algorithm. Additionally, the rejection algorithm significantly boosts precision of final classifier on both coffee and banana datasets, achieving rates of 99.46% and 100.0%, respectively, while using only 50% labeled data. The study also presents a thorough set of experiments and analyses to validate the effectiveness of the proposed method, comparing its performance against state-of-the-art supervised approaches. The results demonstrate that our method, despite using only 50% of the labeled data, achieves competitive performance compared to fully supervised models that use 100% of the labeled data.

Plant phenotyping relevance

葉画像から植物病害状態を分類する半教師あり学習法と曖昧性除去アルゴリズムを開発し、コーヒー・バナナの公開データセットで評価しており、病害表現型の抽出手法が研究の中心である。

Code and data availability

The paper evaluates on two public leaf disease image datasets (BRACOL and a Banana leaf disease dataset from Roboflow), but no author analysis code, trained models, or paper-specific data deposit with an actionable public URL is provided. The Banana dataset's Roboflow link appears only as obfuscated text in the article

No evidence-backed public reproduction asset is currently recorded.