A multispectral imaging analysis for enhancing citrus fruit detection

Abstract

Over the past two decades, a number of researchers around the world tried to develop a citrus harvesting robot. However, no commercial harvesting robot is yet available in the market. Both technical and economic factors have hindered the commercialization of robotic harvesting. Fruit detection in the orchard under natural daylight condition is still a challenging a task. This paper presents a study of using multispectral imaging to enhance citrus fruit detection, in the field under natural daylight condition. The multispectral imaging is composed of a 12-bit monochrome camera fitted with a filter wheel which, carries six optical band pass filters covering the spectrum identified to have a high discriminability between orange fruits and leaves. Multispectral images of mature orange fruit targets were acquired in the field under natural lighting condition. Pattern recognition techniques such as linear discriminant classifier and artificial neural networks were developed to segment the fruit by classifying the fruit pixels from the background pixels. A modified watershed transform combined with blob analysis was used to detect the individual fruits in a cluster. In addition, principal component analysis (PCA) was used to transform the multispectral images and to identify the wavelengths or combination of wavelengths that could improve detection of fruit from the canopy background.