Unified Few Shot Detection and Zero Shot Segmentation Framework for Ripe Strawberry Phenotyping

Abstract

Precision agriculture requires accurate fruit outlining to support automated harvesting and yield assessment. Manual pixel annotation limits scalability and slows deployment in farm environments. An annotation light approach is presented for ripe strawberry detection and instance segmentation across greenhouse and field imagery. The primary claim states reliable masks emerge from coupling few sample trained detectors with prompt driven foundation segmentation. A fast object locator trained with limited images provides region proposals, while a large pretrained segmenter generates masks without pixel supervision. Evaluation uses two datasets with controlled and natural conditions and reports precision recall, intersection over union, and Dice statistics. Results show high detection accuracy under sparse supervision and stable segmentation scores above 0.92 across datasets. These findings advance annotation efficient phenotyping by demonstrating scalability with minimal labeling effort. Applications include real time monitoring, ripeness assessment, and robotic harvesting support. Future work targets multiclass maturity analysis, improved occlusion handling, and multimodal sensing integration.