Automatic detection of orientation and diseases in blueberries using image analysis to improve their postharvest storage quality

The production of the South American blueberry has increased by over 40% in the last decade. However, during storage and shipping, several problems can lead to rejections. This work proposes a pattern recognition method to automatically distinguish stem and calyx ends and detect damaged berries. First, blueberries were imaged under standard conditions to extract color and geometrical features. Second, five algorithms were tested to select the best features to be used in the subsequent evaluation of classification algorithms and cross-validation. The blueberries classes were control, fungally decayed, shriveled, and mechanically damaged. The original 951 features extracted were reduced to 20 or fewer with sequential forward selection. The best classifiers were Support Vector Machine and Linear Discriminant Analysis. Using these classifiers made it possible to successfully distinguish the blueberries' orientation in 96.8% of the cases. By evaluating damages to fungally decayed, shriveled, and mechanically damaged blueberries, the average performances of the classifiers were above 97%, 93.3%, and 86% respectively. All of the experiments were evaluated using external images with 95% confidence - 10-fold cross-validation. These results are promising because they will allow for the increase in export quality when implemented in production lines.

► Blueberry's postharvest three specific diseases were detected using image analysis. ► Blueberry dual-orientation also was detected with performances of 96.8%. ► Detection was implemented using a statistical pattern recognition approach. ► The best classifier performances were above 90% for specific diseases. ► The proposed methodology is promising for the inline sorting blueberries.