Search

Non-destructive quality detection is a major ambition for the development of digital horticulture systems that aims at delivering high-quality fruit to consumers and reduce waste along the food supply chain. X-ray imaging has recently proven to be an accurate technique to determine internal quality of fruit and vegetables, with X-ray Computed Tomography (CT) providing three-dimensional (3D) images and X-ray radiography two-dimensional (2D) images. The latter technique is therefore often regarded as inferior to CT. However, the results demonstrate that a deep learning-based classifier on a single radiograph per apple sample is as accurate as using the full 3D data for fruit leading from hypoxic and anoxic damage. To this end, data on 'Braeburn' apple fruit picked at different orchards and stored under diverse controlled atmosphere (CA) conditions that resulted in internal browning disorders were collected. Various binary 3D and 2D ResNet classifiers were implemented to detect the defect Braeburn apples among the healthy samples, resulting in a model called BraeNet. Performance metrics, such as the accuracy, precision, and recall, together with heatmaps obtained via Guided Grad-CAM were considered to evaluate the model's effectiveness and robustness. The 3D and 2D BraeNet that was trained on healthy apple fruit versus fruit with damage due to storage under N 2 or high CO 2 atmosphere reached an accuracy of 96 ± 1% for both CT and radiography data. The insights obtained in this work will accelerate the development of X-ray detection systems for quality assessment of fruit. [Display omitted] • Deep learning was able to detect defects in X-ray images of apple fruit. • BraeNet, trained on CO 2 damaged fruit, was highly robust to apples stored under N 2. • BraeNet reached an accuracy of 96 ± 1% in detecting defects in X-ray data of apple. [ABSTRACT FROM AUTHOR]