Kvalitetskontroll av potatis : Tranfser learning med Efficientnet-B3

Sammanfattning

Potatoes are susceptible to a range of external damage and diseases that affect both market value and shelf-life. Reliable identification of these defects is essential for efficient production, but manual sorting is time-consuming and inconsistent, while automated systems are often costly. This work investigates how machine learning and image analysis can be used to automate the quality control of potatoes. A convolutional model (EfficientNet-B3) was trained using transfer learning to classify potatoes into five quality classes based on images. The model was first trained with a frozen base and was fine-tuned by unfreezing the final layers. The model achieved an overall classification accuracy of 89,94% on the test set, with the highest performance for the classes green and big, and the lowest performance for the defect class bad. The results show that image-based classification with transfer learning is a promising alternative for automated potato sorting.