Main content area

Combination of shape and X-ray inspection for apple internal quality control: in silico analysis of the methodology based on X-ray computed tomography

van Dael, M., Verboven, P., Zanella, A., Sijbers, J., Nicolai, B.
Postharvest biology and technology 2019 v.148 pp. 218-227
X-radiation, apples, computed tomography, data collection, fruits, image analysis, models, porosity, prediction, quality control, radiography, storage conditions, tissues, vegetables
Multisensor inspection allows for inline detection of internal defects of products with variable shapes, such as fruit and vegetables, by combining X-ray radiography with 3D shape recognition and modelling. For products with a complex internal structure such as apple fruit, inspection must also account for the corresponding internal density gradients that can affect X-ray radiographic images. Apple fruit consist of different tissues with spatially varying differences and gradients in porosity, resulting in corresponding density gradients. Development of internal browning disorders during storage of apples result in damage to tissue structure and associated changes of density distribution. In this paper, a method is developed and evaluated to account for density gradients in apple to improve detection success of defected apples. Experiments were conducted with 26 ‘Braeburn’ apples that were subjected to storage conditions in which internal browning develops. X-ray CT images of each apple were taken at 5 time points over a period of 9 months to obtain a reference dataset of 130 apples with varying levels of defects that could be accurately characterized. The fruit were additionally subjected to multisensory inspection. Using the images of healthy fruit, an internal density distribution model was created and used in the image processing pipeline of the proposed technique to correct for internal gradients. Accounting for a non-uniform internal density distribution increased the R2adj value from 0.73 to 0.86 for prediction of the degree of browning in a fruit. A receiver operator characteristic analysis found that accounting for non-uniform samples reduced the number of false positives from 14 to 5% at a true positive rate of 90%. It was demonstrated that the multisensor inspection approach requires relatively simple hardware of combined with prior knowledge in the form of a statistical shape model and associated density distribution. It is also flexible, since no specific segmentation and detection methods have to be developed for different types of defects.