Deep Learning for Apple Fruit Quality Inspection Using X-Ray Imaging

Astrid Tempelaere, Leen Van Doorselaer, Jiaqi He, Pieter Verboven, Tinne Tuytelaars, Bart Nicolai; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2023, pp. 552-560

Apples are widely consumed worldwide, but the quality of the fruit flesh might deteriorate during storage, resulting in brown tissue formation. X-ray radiography has emerged as a non-destructive method for quickly detecting internal quality problems. This method provides X-ray imaging data that should be processed in an accurate and efficient way. In this paper, we investigate the classification of healthy and defect apples from different orchards and storage conditions using deep learning. The aim of the study was to select a robust and efficient deep learning network that can be used on an X-ray sorting system in a practical setting in the agrifood industry. To this end, the models were evaluated not only in terms of performance but also computational cost. As biological variability is inherent to agrifood problems, we strongly focused on generalizability of the models by using multiple test sets with apples from another orchard and stored under different conditions. The best model had the GoogLeNet architecture, reaching an accuracy of respectively 100 (0)% on a first test set with apples from another orchard, and 82 (8)% on a second test set stored at other conditions. The comparative study provides valuable insights for improving robust and efficient detection algorithms and implementing X-ray technology in the agrifood industry. The proposed technology can be extended to other fruit and vegetables that also suffer from internal quality problems.