From Visual Inspection to Automated Hazelnut Kernel Quality Assessment: A Deep Learning Workflow for Full Cutting-Tray Images

Hazelnut (Corylus avellana L.) kernel quality assessment, a key step in postshelling quality control, is commonly based onvisual inspection of cut kernels, a procedure that can be time-consuming and affected by operator-dependent variability. Thisstudy aimed to develop and evaluate an integrated workflow based on RGB images and deep learning, using a low-complexityacquisition setup, for hazelnut quality assessment from full cutting tray images. The proposed pipeline combined objectdetection, automatic patch extraction, patch-level classification, and tray-level aggregation within a unified computer visionworkflow. Individual hazelnut halves were localized with a YOLOv8s detection model, cropped into image patches and thenclassified by a dedicated classification model as sound, stink bug-damaged or rotten. Model performance was assessed atdetection and patch level, and the workflow was further evaluated at tray-level by comparing aggregated predictions withan operator-derived consensus reference to generate quantitative tray-level quality estimates. The detection stage showedreliable localization of hazelnut halves and provided suitable image patches for classification; the classifier stage performedbetter for sound kernels than for defective classes, with the main uncertainty related to the separation between stink bug-damaged and rotten kernels. Tray-level evaluation further showed that patch predictions could be translated into quantita-tive estimates of sample quality composition, although class-specific deviations remained for visually overlapping defects.Overall, the study demonstrates that RGB image-based deep learning can provide structured and reproducible support forpostshelling hazelnut quality control, integrating object localization, kernel classification, and tray-level quality estimationwithin a single automated workflow. (PDF) From Visual Inspection to Automated Hazelnut Kernel Quality Assessment: A Deep Learning Workflow for Full Cutting-Tray Images. Available from: https://www.researchgate.net/publication/417607873_From_Visual_Inspection_to_Automated_Hazelnut_Kernel_Quality_Assessment_A_Deep_Learning_Workflow_for_Full_Cutting-Tray_Images [accessed Nov 11 2026].