Answers are grounded in this paper and cite their sources.
Counting grains on a rice panicle by threshing is time-consuming, labour-intensive and destructive to panicle architecture, whereas counting them in situ is hampered by the slender, randomly oriented and densely packed habit of the grains, which horizontal bounding boxes (HBBs) describe poorly. We therefore propose SDP-YOLO, an oriented bounding box (OBB) detector for on-panicle grain counting, and the remote pipeline in which it is evaluated. Panicles are excised at maturity and imaged whole on a laboratory platform, with no threshing or grain separation, so that the sampling is destructive to the plant but preserves the panicle architecture and grain arrangement that threshing destroys. On an OBB-annotated dataset of 321 images, SDP-YOLO inserts SPDConv in the backbone to suppress detail loss during downsampling, DySample in the neck to correct upsampling misalignment, and a P2 detection head that provides a 160 × 160 prediction surface for small, densely packed grains. Under five-fold cross-validation, which is used as the primary basis for comparison, SDP-YOLO reaches 95.96 ± 0.35% mAP@0.5 and 70.12 ± 0.66% mAP@0.5:0.95 against 90.03 ± 0.84% and 42.85 ± 2.13% for YOLO11n-OBB, a gain of 5.93 and 27.27 percentage points with a smaller standard deviation on every metric. Four random seeds and an independent 60-image test set covering additional varieties reproduce the ranking; on that set, the counting mean absolute error (MAE) falls from 8.90 to 3.97 grains per panicle and R2 rises from 0.8957 to 0.9540. An edge–cloud system, in which a Raspberry Pi 5 acquires images and a cloud server performs inference, completes the end-to-end loop in 1956.55 ms on average. OBB annotation combined with detail-preserving downsampling, content-aware upsampling and a high-resolution head thus provides an accurate and deployable route to on-panicle rice grain counting.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: