Generative Monocular Perception Pipeline-Based Framework for Accurate Stem Detection in Automated Strawberry Harvest

Horticulture faces a growing labor crisis, driving demand for autonomous harvesting robots, but reliable strawberry peduncle detection remains a critical unsolved challenge due to their fine, millimeter-scale structure and severe intertwining with leaves and stems. Existing single-view RGB imaging struggles with occlusions and ambiguities, while depth sensors falter in reflective greenhouse environments plagued by noise and data gaps. Introducing a generative monocular perception pipeline—the first to reconstruct multi-view cues purely from a single RGB image—this study achieves perceptual consistency through four novel, synergistic innovations: (i) pseudo multi-view synthesis to emulate diverse viewpoints, (ii) monocular depth estimation for precise geometric guidance and background isolation, (iii) line-curve geometric modeling to capture subtle peduncle features, and (iv) occlusion-order reasoning via cross-view consistency analysis. In comparative trials against a YOLO-based detector (85.71% region accuracy vs. 57.14%), our pipeline delivers orientation precision, slashing mean angular error from 18.31° to 13.96°—robust, clutter-resilient cutting cues for next-generation robotic harvesters. Evaluated on farm images, it reduces mean angular error to 13.96° (SD 10.15°) from YOLO's 18.31° (SD 11.27°), with p<0.05 (paired t-test, n=14).

@article{Arai2026,
  title     = {Generative Monocular Perception Pipeline-Based Framework for Accurate Stem Detection in Automated Strawberry Harvest},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {17},
  number    = {4},
  year      = {2026},
  publisher = {The Science and Information Organization},
  author    = {Kohei Arai and Jin Sawada and Mariko Oda},
  doi       = {10.14569/IJACSA.2026.0170414},
  url       = {https://doi.org/10.14569/IJACSA.2026.0170414}
}