SSPM-YOLOv11 and SSM-3D: Efficient oriented object detection and sub-millimeter-accurate morphometric measurement for morels (Morchella spp.) in agricultural grading scenarios

• Proposed a lightweight model (SSPM-YOLOv11) for oriented morel detection. • Achieves faster and more accurate detection compared to the original model • SSM-3D enables non-contact morphometric measurement with sub-millimeter accuracy • Integrated detection and measurement supports automated harvesting and grading • Stable detection and morphometric measurement under complex grading scenarios In agricultural grading scenarios, morel ( Morchella spp.) targets typically exhibit pronounced morphological variability, arbitrary orientations, frequent occlusions, dense distributions, and complex background interference arising from changes in ambient illumination conditions and imaging noise. These factors limit the performance of manual grading and conventional image-processing methods in terms of detection accuracy and morphometric measurement robustness. To address the aforementioned challenges, we propose an OBB detector for morels, termed SSPM-YOLOv11, and a sub-millimeter morphometric measurement method, SSM-3D. SSPM-YOLOv11 is built on YOLOv11-OBB and integrates MSA-ARConv to construct a multi-scale adaptive receptive field, RIAttention to enhance rotation-invariant representations, and C3k2-WT to improve texture and structural feature extraction. For morphometric measurement, an Intel RealSense D435i depth camera is employed, and SSM-3D treats the midpoints of the short opposite sides of each predicted OBB as keypoints and projects them into 3D space. Euclidean distances between paired 3D keypoints are computed to obtain the morel cap length and diameter. Experiments on the self-constructed “Seven-Sister Morel” dataset show that, relative to YOLOv11n-OBB, SSPM-YOLOv11 improves mAP@0.5 and mAP@0.5:0.95 by 7.5 and 9.8 percentage points, respectively, and boosts recall by 6.1 percentage points. In addition, FLOPs decrease by 4.2%, and inference runs at 186 FPS. For morphometric measurement, predictions closely match manual measurements, with AE within 1 mm for most samples and an MAE of 0.83 mm. Overall, the proposed framework provides efficient and reliable visual perception for automated morel harvesting and grading, and offers transferable insights for detecting and quantifying non-rigid, oriented objects.