AUTHOR=Liu Bo , Wei Shusen , Zhang Fan , Guo Nawei , Fan Hongyu , Yao Wei TITLE=Tomato leaf disease recognition based on multi-task distillation learning JOURNAL=Frontiers in Plant Science VOLUME=Volume 14 - 2023 YEAR=2024 URL=https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2023.1330527 DOI=10.3389/fpls.2023.1330527 ISSN=1664-462X ABSTRACT=Tomato leaf diseases can cause major reductions in both crop yield and quality. Automated recognition of leaf diseases using computer vision techniques has shown promise but faces challenges including variations in symptoms, limited labeled data, and model complexity. While prior works have explored hand-crafted features and deep learning models for both tomato disease classification and multi-task learning with severity estimation, little attention has been paid to sufficiently exploiting the shared and unique knowledge between these two tasks. This paper presents a novel multi-task distillation learning (MTDL) framework aimed at comprehensive diagnosis of tomato leaf diseases. MTDL employs a multi-stage learning strategy that includes knowledge disentanglement, mutual learning, and knowledge integration to effectively utilize the complementary characteristics of disease classification and severity prediction tasks. To further simplify the training process and reduce dependency on traditional teacher models, we introduce a decoupled teacher-free knowledge distillation (DTF-KD) strategy that employs a virtual teacher for more adaptive learning. Experimental results reveal that our framework enhances performance while reducing model complexity. Our MTDL-optimized EfficientNet outperforms the single-task ResNet101 baseline in disease classification accuracy by 0.68% and in severity estimation by 1.52%, while requiring only 9.46% of its parameters. These findings highlight the framework's potential for practical application in intelligent agriculture.