Shiitake mushroom cap-stem adaptive matching and harvesting point localization for edge devices

Abstract

Robotic shiitake harvesting requires reliable target perception, individual-level structural association, physical measurement, and actionable picking-point localization under edge-computing constraints. To address these requirements, this study developed a color-and-depth edge perception framework for shiitake harvesting. A lightweight instance segmentation model was constructed using a Mushroom Structure-Aware Attention module and a Background Suppression Attention module to enhance heterogeneous cap-stem features and suppress interference from mushroom-stick backgrounds. A scale-adaptive cap-stem association method was proposed by integrating direction-specific two-dimensional spatial constraints with physical-scale-dependent three-dimensional depth constraints. This method converts perception results into physical cap diameters and three-dimensional stem picking points. The proposed model achieved a mask mean average precision of 53.2% across intersection-over-union thresholds from 0.50 to 0.95, representing an improvement of 3.0 percentage points over the baseline model, and maintained low computational complexity. The proposed cap-stem association method achieved an F1-score of 97.89%, and single-frame inference on the Jetson Orin Nano required 21.41 ms. Cap diameter estimation achieved a coefficient of determination of 0.974 and a root mean square error of 2.093 mm at an observation distance of 250 mm. These results indicate that the proposed framework can provide real-time and physically interpretable perception outputs for subsequent robotic manipulation and selective harvesting.