Blueberry fruit detection by Bayesian classifier and support vector machine based on visible to near-infrared multispectral imaging

Early yield estimation based on computer vision enables better labor deployment and lower harvest expense in a large scale blueberry field. In this study, ninety multispectral images with near-infrared (NIR), red(R) and green(G) bands were collected from southern Highbush blueberry variety 'Sweetcrisp' from a commercial blueberry field in Waldo, Florida, during 20 April 2011 and 15 May 2011. Five thousand pure fruit pixels and 5000 background pixels were collected from the images. 66% of them were in a calibration set and the other 34% were used as a validation set. Various representations of the multispectral color models (MHSI, MYIQ, MYCbCr) originated from the NIR-R-G color model were used as the features. Bayesian classifier and support vector machine were applied for the classification of the fruit and background classes. Principle component analysis was applied before Bayesian classification for the optimized use of the features. Results show that support vector machine outperformed the Bayesian classifier with higher true positive rate (84% for fruit class and 73% for background class) and lower false positive rate (27% for fruit class and 16% for background class) in the fruit/background classification. In addition, 1000 pixels of each of eight classes (mature fruit, mid-mature fruit, young fruit, leaf, branch, soil, sky and reference board, which were found in most images) were also classified by using the two classification techniques. The true positive rates for mid-mature fruit and young fruit class were around 50%, which indicates that the color spaces were not useful for the classification of different fruit stages.