ALDNet: A two-stage method with deep aggregation and multi-scale fusion for apple leaf disease spot segmentation

Deep learning has been successfully applied in plant leaf disease segmentation in simple environments. However, complex environments pose challenges like blurred boundaries and small disease spots. To tackle these issues, this paper proposes ALDNet, a two-stage method utilizing PBGNet for leaf segmentation and PDFNet for disease spot extraction. PBGNet incorporates a deep aggregation (DA) module to generate an initial segmentation map enriched with global contextual information, and the features guided by the reverse attention (RA) are fused to the shallowest levels to enhance target leaf boundary precision. PDFNet combines residual path and multi-scale fusion (MSF) to enhance feature representation and fuse semantic and scale features from different branches, facilitating small spot extraction. ALDNet outperforms state-of-the-art methods in complex environments, achieving 94.53% IoU and 98.13% PA for leaf segmentation, and 77.41% mIoU and 84.1% mPA for disease segmentation. ALDNet provides a reliable solution for accurate severity grading of apple leaf diseases.