Leveraging grounded SAM and a weakly supervised CNN filtering for enhanced leaf segmentation in automated plant phenotyping

This study presents a novel framework for leaf segmentation, integrating Grounded SAM with a lightweight CNN to enhance automated plant phenotyping. By leveraging prompt-derived bounding boxes, Grounded SAM provides initial segmentation, which is refined by a CNN trained on minimal data to isolate leaf segments. The method was validated on potato and CVPPP datasets, demonstrating improved recall and precision over the baseline. This approach reduces reliance on extensive training data, offering a scalable solution for diverse plant phenotyping applications.