Apple Leaf Disease Segmentation in the Wild: A Multi-task Collaborative Learning Approach

Bo Liu

Abstract: Apples are economically significant but threatened by leaf diseases. While deep learning methods have advanced disease segmentation, these methods struggle with real-world challenges like complex backgrounds, overlapping leaves, and varying lighting. Considering the scale and visual differences between leaf and disease regions, most existing approaches adopt a two-stage strategy: leaf localization followed by disease segmentation within the leaf. However, errors in the leaf localization stage can negatively impact the disease segmentation stage, leading to imprecise segmentation and computational inefficiency. To address these limitations, we propose LDS-MTCL, a novel multi-task collaborative learning framework for leaf disease segmentation, which unifies leaf localization and disease area identification tasks. LDS-MTCL integrates two key modules: leaf area localization (LAL), which employs a weighted pyramid aggregation (WPA) operation to accurately define leaf regions, and disease area enhancement (DAE), which performs fine-grained disease segmentation guided by LAL. The architecture of LDS-MTCL facilitates the exchange of learned features between the two modules, enabling the model to better understand the spatial context of leaf and disease regions. Extensive experiments demonstrate LDS-MTCL outperforms existing two-stage models in terms of segmentation performance and efficiency. Specifically, our approach achieves a mean intersection over union (mIoU) of 93.68% and a mean pixel accuracy (mPA) of 96.43%, surpassing advanced techniques by 1.75% in mIoU and 2.12% in mPA, while significantly reducing inference time by up to 81.31%. This work contributes to the advancement of intelligent agriculture systems.