Cimiciato defect detection in hazelnuts: CNN models applied on X-ray images

Hazelnuts are a significant crop with an increasing importance, especially for confectionery industry. Insect damages affect hazelnut quality, requiring post-harvest selection based on industrial quality standards which often exceed official regulations. Currently used methods for identifying insect damages (cimiciato) often rely on visual inspection, external imaging or require destructive testing. This study compared twelve different pretrained Convolutional Neural Network (CNN) architectures applied on hazelnut kernels X-ray radiographs for the automated detection of cimiciato defects. Through an extensive training and validation process, followed by testing on a separate dataset, InceptionV3 architecture showed the best overall balance across all performance metrics, including accuracy, sensitivity, and precision, while Xception demonstrated superior specificity and the lowest false positive rate. Lightweight models such as SqueezeNet and ShuffleNet provided fast and resource-efficient training, though with moderate trade-offs in classification accuracy. In contrast, deeper architectures like Inception-ResNet-V2 and Xception, while computationally demanding, achieved greater robustness and generalization capability. Our findings suggest that some CNN architectures combined with X-ray imaging could effectively be employed in a reliable and efficient non-destructive selection method for the hazelnut industry, potentially improving product quality control and minimizing losses associated with insect damages. • X-ray imaging of hazelnut kernels allows to visualize all insect damages • Pretrained CNNs applied to X-ray images allows to classify healthy VS damaged hazelnuts • InceptionV3 and Xception were the most performing CNN architectures • CNN with X-ray imaging is a promising approach for non-destructive sorting of hazelnuts