Public summary

Intermediate validation test deep learning

Greefa

February 27, 2025

Apple quality sorting based on deep learning

Sorting apples for quality is a complex process that requires accurate classification of defects. Greefa conducted a validation test to evaluate the applicability of deep learning in automatically recognizing and distinguishing defects. By using artificial intelligence, the sorting process can be performed more efficiently and accurately, with less manual control.

Innovation package, use case and type test

    Status: completed

    Technical functionality

    TF-1

    Broad knowledge question

    How accurately can deep learning classify fruit?

    This research focuses on the extent to which deep learning is suitable for classifying fruit and distinguishing different defects. The test assesses whether this technology can contribute to a more stable and efficient sorting process.

    Approach

    Testing deep learning in a realistic environment

    The validation test took place on a sorting machine at FruitMasters in Geldermalsen. For this test, 1,400 Kanzi apples were selected by two inspectors to create a representative set of possible defects. The apples were then sorted by type, with each defect being assigned to a specific class. The test results were recorded and compared both manually and by the sorting system.

    Goal

    Comparison with existing sorting methods

    The goal of this validation test is to evaluate the deep learning method in a controlled but realistic setting. In addition, the new method is compared to existing sorting techniques to determine whether deep learning provides improved accuracy and efficiency.

    Result and reflection

    Better detection and more stable sorting results

    The deep learning method proves highly accurate in distinguishing apples with and without defects.

    Successes

    • The system achieved high specificity (>99%) and sensitivity (>94%) in defect classification.

    • Deep learning can better differentiate between various defect types than traditional sorting methods.

    • The method works independently of apple variety, enabling broader applicability.

    Lessons learned

    • Test results, based on a wide range of defects, show low variation (0.5%–2.5%).

    • Repeating the test with the same apple variety is not expected to produce significant differences.

    • Less manual reinspection is needed, potentially leading to a more efficient and labor-saving sorting process.

    Leading Partners involved