Small-Lesion and Boundary-Aware Mask2Former for Pixel-Level Segmentation and Severity Assessment of Cucumber Target Spot Disease

Cucumber target spot disease, caused by Corynespora cassiicola, is a major constraint on greenhouse cucumber production. Accurate pixel-level segmentation and automated severity assessment remain challenging because early lesions are small, lesion boundaries are gradual, and field images often contain complex illumination and background interference. To address these issues, this study proposes a small-lesion and boundary-aware Mask2Former framework with a ResNet-50 (R50) backbone for cucumber target spot segmentation. Specifically, a Small Target Enhancement Module (ST) is designed by integrating atrous spatial pyramid pooling (ASPP), a high-resolution feature retention branch, and a small-target-weighted loss to improve sensitivity to early micro-lesions. In addition, a Boundary Segmentation Module (BS) is introduced to enhance boundary localization through boundary attention and explicit supervision with Dice and Focal losses. A field dataset containing 2559 cucumber leaf images was annotated at the pixel level and split into training, validation, and test sets at a ratio of 7:1:2. Disease severity was categorized into four grades based on the lesion-to-leaf-area ratio. On the test set, the proposed model achieved an mIoU of 85.05%, a Dice coefficient of 91.94%, and a precision of 92.30%, outperforming the Mask2Former baseline by 4.95, 4.72, and 4.58 percentage points, respectively. Moreover, severity assessment based on segmentation results reached an overall grading accuracy of 91.6% (Cohen’s Kappa = 0.89), with misclassifications predominantly confined to adjacent severity grades and no large cross-category errors observed. These results indicate that the proposed method has practical potential for automated disease monitoring and precision management in greenhouse cucumber production.