Automated detection of internal fruit rot in Hass avocado via deep learning-based semantic segmentation of X-ray images

Abstract

Internal rot of avocado fruit (Persea americana), attributable to fungal infection, occurs at the end of the ripening process and causes only minor changes in the appearance and texture of the fruit surface. Manual inspection of rot by sight and touch commonly conducted in countries importing avocado fruit is time-consuming, labor-intensive, and subjective. In this context, X-ray line scanning has been proven as an advantageous method of fruit rot detection because of its speed of data acquisition and the indication of internal rot by bright regions in associated images. However, some fruit internal disorders exhibit only poor changes in contrast, resulting in low detectability by traditional image processing. This study aimed to test the effectiveness of a detection model using deep learning-based semantic segmentation in identifying two types of fruit rot, stem-end and body rot, in Hass avocados. Therefore, U-net+ + was trained and validated via 5-fold cross-validation to classify every pixel in an X-ray image as either infected or not. Then, each X-ray image was binarily classified based on either the presence or absence of internal fruit rots, achieving an accuracy of 0.98. Furthermore, the percentage of infected area was quantified with a root mean squared error (RMSE) of 3.15 %. Lastly, the proposed model detected both stem-end and body rot as well as rot along low-contrast fruit edges. The results of this study indicate that the proposed automatic inspection system using deep learning-based X-ray image analysis can effectively detect internal rot in Hass avocado fruit. This non-destructive, objective detection model can therefore increase efficiency and reduce misclassification in post-harvest avocado inspection. Furthermore, deep learning-based X-ray imaging has potential for applications in fruit inspection for internal cavities attributable to diseases or wounds.