Olive oil content prediction models based on image processing Academic Article uri icon


  • An increase in olive oil content in harvested olives can be achieved by optimizing the harvest time. In order to determine optimal harvest time, prediction models were developed to determine oil content based on quality features derived from known image processing algorithms. Digital colour photographs were taken weekly of opposite sides of large samples of Picual and Souri olives during the ripening season. In addition, the fallout was collected and weighed weekly during the same period. Quality features such as size, shape, colour, and texture were derived from the photographic images of the olives. Low resolution nuclear magnetic resonance was used for rapid and accurate determination of the oil content of each photographed olive. The correlations between the various quality features and oil content were determined. Two prediction models based on linear regressions and artificial neural networks were developed to predict the quantity of oil. The sensitivity of the models to the quality characteristics, different proportions between training and testing sets, the topologies of the networks, and various transfer functions were analyzed. Based on factor analysis, the characteristics that most strongly affect oil quantity for both varieties were found to be colour, texture, size, and shape. The neural network models were more accurate than linear regression, resulting in average linear correlations of 0.81 and 0.87 to the oil quantity in Picual and Souri olives, respectively. Improved results were obtained when features from images of both sides of the olives were considered. It is recommended to develop a specific model for each olive variety.

publication date

  • February 1, 2010