Estimating strawberry size using improved YOLOv8s-seg and semi-truncated cone models

Accurate and real-time three-dimensional size estimation is often required for strawberries under agricultural field environments. However, the vision systems are typically restricted to the data capture from a single, oblique viewpoint. Only a partial three-dimensional point cloud can be generated with the practical constraints in the ridge-planting systems. The visible upper surface of the fruit can be represented without the full geometric information. Concurrently, a fundamental modeling challenge can also exist, due mainly to the natural morphology of a strawberry. A conical shape can be characterized by a truncated calyx base and a tapered tip, which is deviated substantially from the simple symmetric shapes, such as the ellipsoids commonly used in the conventional fitting. This geometric mismatch has been one of the primary sources of the estimation inaccuracies. Particularly, the critical dimensions (like the bottom diameter) cannot be directly observed from the available data perspective. In this study, an integrated computational framework was proposed to combine a lightweight and enhanced segmentation network with a purpose-specific geometric model designed to match the true morphology of the fruit. A four-stage pipeline was constructed for data completeness and model accuracy. 1) The robust instance segmentation was first performed by an improved YOLOv8s-seg network. The backbone was augmented with a custom C2f_Faster_EMA module in order to enhance the multi-scale feature extraction for the fine details with minimal computational overhead. 2) The high-precision mask was fused with the synchronized depth data to generate an initial 3D point cloud. The critical preprocessing involved the statistical outlier removal and voxel grid downsampling. Several thousand raw points were efficiently condensed into a consistent set of approximately 100 to 150 representative points, thereby preserving the key geometric features to reduce the computational load. 3) Subsequently, the hybrid RANSAC-PCA algorithm was operated on the refined point cloud to robustly estimate the strawberry's principal axis and 3D centroid. 4) The final stage employed a semi-truncated cone model, which was defined by the height, top diameter, and bottom diameter. The processed point cloud was used to fit the size estimation after optimization. The performance of the framework showed that the segmentation model was achieved in a mAP@0.5 of 96.9% and a ripe fruit segmentation accuracy of 98.3%, while there was a high inference speed of 159.3 frames per second. The size estimation shared high accuracy under varied conditions. Tests showed that the average relative errors of 1.2%, 0.7%, and 1.7% were obtained for the longitudinal diameter, top diameter, and bottom diameter, respectively, under different viewing angles. The relative errors of 1.8%, 1.2%, and 4.8% were found over a practical range of the capture heights from 30 to 70 cm after evaluations, indicating the consistent robustness. A multi-sample test was also performed on 22 standard strawberry specimens. The precision was further validated by the average value of errors of -0.13, 0.03, and -0.03 mm for the three dimensions, respectively, indicating all within the sub-millimeter range. Furthermore, the direct comparative analysis showed that the improved model significantly outperformed the baseline approaches. The absolute estimation error was also reduced by 50% and 70%, respectively, for the top and bottom diameters, compared with a single-view TSDF method. Crucially, an accurate estimate was successfully provided for the bottom diameter, rather than in the semi-elliptical models. The end-to-end pipeline also required an average processing time of 3.3s per frame, indicating the feasibility and strong potential. The real-time field deployment can be extended to more applications, such as the strawberry phenotyping, in-field grading, and precision yield prediction. The finding can also offer a practical solution using affordable RGB-D sensing hardware.