Axial Orientation Recognition and Two-Stage Rotational Correction of Cucurbit Rootstock Seedlings

To address the difficulty of stable axial orientation determination of cucurbit rootstock seedlings under random initial postures, true-leaf interference, and mechanical occlusion during mechanized grafting, a two-stage axial orientation recognition and rotational correction method integrating mechanical constraint and machine vision was proposed. In the first stage, seedling posture was normalized using the PCA axis of the whole green region, and the seedling was rotated to avoid the leaf hold-down ring handle. In the second stage, mechanically occluded leaf regions were restored after ring constraint, and a minimum-span half-area window method based on polar-angle distribution was proposed to determine the directed cotyledon axis. The method was compared with the rotation-center–cotyledon-centroid line method and the single-cotyledon PCA method. Fifty seedlings were tested at each of the true-leaf emergence, initial true-leaf expansion, and true-leaf extension stages. In 150 first-stage rotation trials conducted on 50 seedlings, the absolute PCA-axis deviation after rotation was within 1.5° in every trial. In a separate handle-avoidance test, the avoidance rate was 100%. In the second stage, the proposed method yielded median absolute errors of 0.71°–1.26° for both cotyledons, with no significant differences among growth stages. In contrast, the errors of the centroid-line and PCA methods increased significantly with seedling development, reaching maximum median errors of 5.98° and 19.22°, respectively. After the second-stage rotation, the algorithm-derived residual axial deviation of the right cotyledon was below 1° for all 150 seedlings. These results indicate that the proposed method provides stable cotyledon-axis recognition under the investigated experimental conditions.