Integration of multi-scale EMA and deformable high-frequency enhancement for instance segmentation and disease severity assessment in pear black spot disease

• Improved EF-Mask2Former integrates EMA, HF-EMA and FasterNet modules for lesion segmentation. • EF-Mask2Former with multi-scale EMA & deformable HF-EMA solves pear black spot segmentation challenges. • FasterNet-based feature decoupling preserves micro-lesions;Integrated rchitecture enables end-to-end lesion segmentation to severity grading. • Achieves 93.2% mAP75 (+12.9% gain) and 9.6% higher micro-lesion detection via deformable edge enhancement. Effective early prevention and management of pear black spot disease are crucial for minimizing losses in pear trees and maintaining agricultural productivity. To address challenges such as segmentation of pear black spot disease, which includes blurred edges, missed slight detection, and poor adaptability to complex lighting conditions, this study proposes an improved Mask2Former model that integrates multiscale fusion and high-frequency edge enhancement. The model establishes a collaborative optimization framework of “global perception detail preservation edge reconstruction”, achieving end-to-end optimization from small target lesion segmentation to disease severity assessment. The enhanced architecture incorporates an efficient multi-scale Attention (EMA) module, constructing dual-branch parallel structures in deep backbone layers. This design enhances global semantic robustness and enables adaptive receptive field adjustment, thereby precisely capturing the multi scale morphological features of black spot disease. Additionally, a cross-layer partial convolution mechanism is employed to dynamically decouple pathological features from background interference, effectively addressing the texture loss of early micro-lesions caused by repeated downsampling. Furthermore,the Deformable High-Frequency Enhancement Module (HF-EMA) integrates orientable Sobel filtering with deformable convolutional kernel adaptation. This integration facilitates the subtle edge reconstruction at the boundaries between diseased and healthy tissues. Experimental results demonstrate that the improved model achieves a segmentation accuracy of 93.18% (mAP 75 ) in open field orchard environments, representing a 12.93% improvement over the baseline, with a 9.56% increase in micro-lesion detection accuracy. The multi-dimensional severity assessment system, based on the segmentation results, achieves classification accuracies of 82.99%, 83.81%, and 89.57% for slight, medium, and serious infection levels, respectively, when using the lesion area proportion criteria. This model provides a high-precision and robust solution for intelligent orchard disease monitoring. Meanwhile, its dynamic feature decoupling mechanism and bio-inspired edge enhancement strategy establish a technical framework for pathological quantification in complex agricultural environments.