CMSAF-Net: integrative network design with enhanced decoder for precision segmentation of pear leaf diseases.

Abstract

Pear leaf diseases represent one of the major challenges in agriculture, significantly affecting fruit quality and reducing overall yield. With the advancement of precision agriculture, accurate identification and segmentation of diseased areas are critical for targeted disease management and optimizing crop production. To address these issues, this study proposes a novel segmentation model, CMSAF-Net, for pear leaf diseases. CMSAF-Net integrates a Multi-scale Convolutional Attention Module (MBCA), a Self-adaptive Attention-augmented Upsampling Module (SAUP), and a Cross-layer Feature Alignment Module (CGAG) to enhance feature extraction, preserve edge information in complex disease regions, and optimize cross-layer information fusion. Additionally, CMSAF-Net incorporates pre-trained weights to leverage prior knowledge, accelerating convergence and improving segmentation accuracy. On a self-constructed dataset containing three types of pear leaf diseases, experimental results demonstrate that CMSAF-Net achieves 88.65%, 93.36%, and 93.86% in key metrics of MIoU, MPA, and Dice, respectively. Compared with mainstream models such as Unet++, DeepLabv3+, U $_{}^2$ -Net, and TransUNet, CMSAF-Net exhibits significant performance improvements, with MIoU increases of 2.45%, 3.86%, 2.21%, and 8.28%, respectively. This study highlights CMSAF-Net's potential for large-scale disease monitoring in intelligent agriculture, providing an efficient segmentation solution with substantial theoretical and practical implications.