An improved DeepLabV3+ based approach for disease spot segmentation on apple leaves

Abstract

This study presents an improved DeepLabV3+ model named AS-DeepLabV3+, specifically designed for segmenting disease spots on apple leaves. AS-DeepLabV3+ addresses critical challenges such as blurry spot edges, the small proportion of spot pixels, and significant variation in spot shapes. First, MobileNetV2 is employed as a lightweight backbone, reducing model complexity. Second, multiple attention mechanisms—Coordinate Attention, ECA Attention, CBAM, and Triplet Attention—are integrated into a unified Multi-Attention module, enhancing feature representation. Third, a dynamic Atrous Spatial Pyramid Pooling (ASPP) module is introduced to effectively capture multi-scale features. Lastly, dense connectivity is utilized in the decoder to improve feature reuse and detail recovery. The model was trained and validated on a dataset of 6,400 apple leaf images collected under natural lighting conditions. Experimental results demonstrate that our proposed model achieves a mean Intersection over Union (mIoU) of 98.00 %, a mean Pixel Accuracy (mPA) of 98.95 %, and a precision of 98.45 %, significantly outperforming existing models, including the original DeepLabV3+, SegNet, BiSeNet, PSPNet, and U-Net. Furthermore, the model has been integrated into a WeChat Mini Program to offer efficient and reliable disease detection services for agricultural practitioners.

Plant phenotyping relevance

リンゴ葉の病斑を画像から分割・検出する手法の開発と検証が研究の中心であり、植物の病害状態を直接推定しているため。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.