Nondestructive detection of the external dimensions and internal quality of Kashi prunes based on YOLO-HSI

Aiming at the inherent limitations of traditional single-point spectral detection methods in terms of simultaneously achieving population target localization and multi-dimensional quality characterization, as well as the significant engineering challenge where the dynamic extraction of spectral Region of Interest (ROI) is highly susceptible to background noise interference in unstructured industrial environments, a comprehensive multi-dimensional quality detection and commodity grading framework (abbreviated as YOLO-HSI) based on the synergistic perception of an oriented object detection network (YOLO-OBB) and push-broom hyperspectral imaging (HSI) was proposed for Kashi prunes to achieve rapid non-destructive inspection.By means of introducing the deep-learning-based YOLO-OBB network combined with a newly developed geometric homomorphic contraction strategy, this study successfully overcame the technical bottlenecks concerning high-precision spatial localization for fruits with arbitrary random tracking postures and the dynamic extraction of pure, unpolluted spectral profiles. On the basis of this spatial-spectral synergy, a comprehensive evaluation system encompassing external dimension measurement, internal soluble solids content (SSC) quantitative inversion, and early latent moldy defect elimination was effectively constructed. To implement this proposed quality assessment framework, consecutive data optimization involving Savitzky-Golay smoothing, Min-Max normalization, and second-derivative transformation was rigorously conducted on the extracted pure ROI spectra, followed by feature variables selection and the formulation of a multi-indicator comprehensive grading rule according to national fruit quality guidelines. The comprehensive experimental results demonstrated that the optimized YOLOv11n-OBB model achieved outstanding localization and detection capabilities, yielding high detection recall rates of 0.992 for intact prunes and 0.982 for moldy prunes, while achieving ultra-fast online processing speeds of 243 frames per second and 270 frames per second, respectively, which reflects a significant advantage for real-time high-throughput inspection. Based on the pixel-to-physical world coordinate mapping relationship, automated millimeter-level measurements of fruit polar diameter and equatorial diameter were successfully realized, where the size prediction root mean square errors (RMSE) fell below 1.1 mm for intact prunes and 0.8 mm for moldy prunes. In terms of internal quality detection, the classification model based on second-derivative preprocessing achieved a complete discrimination accuracy of maximum and an ideal area under the curve (AUC) score of 1 for identifying early latent moldy fruits, while successfully compressing spectral variables to achieve a feature wavelength reduction rate of 70.1%. Furthermore, the CARS-Ridge multivariate regression model exhibited robust predictive performance for internal soluble solids content, yielding a high prediction coefficient of determination (R²_p) of 0.7225 and a reliable ratio of performance to deviation (RPD) of 1.90. The multi-indicator grading decision framework demonstrated that the comprehensive commodity grade determination accuracy reached 85.00% for premium, first-grade, and second-grade prunes, with cross-grade misclassifications occurring exclusively between adjacent boundary levels without severe grade-skipping deviations. In conclusion, the findings of this study successfully validate the high effectiveness and feasibility of the image-spectrum synergistic technical route for the non-destructive detection of multi-dimensional quality in Kashi prunes. The excellent detection performance and classification accuracy achieved by the proposed framework can provide core algorithms and data-driven theoretical support for the research, development, and industrial edge deployment of a new generation of intelligent fruit sorting equipment.