Segmentation-Based Detection for Luffa Seedling Grading Using the Seg-FL Model

This study addresses the issue of inaccurate and error-prone grading judgments in luffa plug seedlings. A new Seg-FL seedling segmentation model is proposed as an extension of the YOLOv5s-Seg model. The small leaves of early-stage luffa seedlings are liable to be mistaken for impurities in the plug trays. To address this issue, cross-scale connections and weighted feature fusion are introduced in order to integrate feature information from different levels, thereby improving the recognition and segmentation accuracy of seedlings or details by refining the PANet structure. To address the ambiguity of seedling edge information during segmentation, an efficient channel attention module is incorporated to enhance the network’s focus on seedling edge information and suppress irrelevant features, thus sharpening the model’s focus on luffa seedlings. By optimizing the CIoU loss function, the calculation of overlapping areas, center point distances, and aspect ratios between predicted and ground truth boxes is preserved, thereby accelerating the convergence process and reducing the computational resource requirements on edge devices. The experimental results demonstrate that the proposed model attains a mean average precision of 97.03% on a self-compiled luffa plug seedling dataset, representing a 6.23 percentage point improvement over the original YOLOv5s-Seg. Furthermore, compared to the YOLACT++, FCN, and Mask R-CNN segmentation models, the improved model displays increases in [email protected] of 12.93%, 13.73%, and 10.53%, respectively, and improvements in precision of 15.73%, 16.93%, and 13.33%, respectively. This research not only validates the viability of the enhanced model for luffa seedling grading but also provides tangible technical support for the automation of grading in agricultural production.