FLAsH: FastSAM-based lettuce analysis for high-density cultivation

초록

With the advent of computer vision, various studies have been conducted on analyzing lettuce growth in smart farming. However, most existing methods assume sparse, single-plant settings and consequently struggle to handle the dense canopy overlap and limited annotation resources prevalent in real-world smart farms. To overcome these limitations, we propose FLAsH, a FastSAM-based lettuce analysis framework for high-density cultivation that jointly addresses instance segmentation and individual growth prediction in a single end-to-end pipeline. While vanilla FastSAM tends to over-segment individual leaves rather than recognize each plant as a unified entity, we fine-tune FastSAM on smart farm images to address this limitation and enable accurate plant-level instance segmentation under limited data conditions. For each detected plant, we use CNN models to predict days after transplanting, current weight, and harvest weight, enabling both current growth analysis and harvest stage prediction. Experimental results demonstrate that the proposed framework achieves an average mAP of 0.788 across six datasets in the segmentation stage and an average MAPE below 10% across all tasks in the growth prediction stage, confirming the accuracy and robustness of the proposed method. This study establishes a practical framework for automated lettuce monitoring in high-density smart farms, enabling growth assessment and harvest planning without manual plant separation or destructive sampling. The code is publicly available at https://github.com/hyeseunng/FLAsH.

키워드