Citrus Segmentation for Automatic Harvester Combined with AdaBoost Classifier and Leung-Malik Filter Bank

Abstract

Image segmentation was the pre-processing step of fruit detection and positioning, and was challenging due to illumination change, occlusion, and cluttered background. To address segmentation problem, a novel segmentation method using AdaBoost classifier and texture-colour features was presented. The proposed method included two steps: (1) a fixed size sub-window was slid along every site in the image, and colour and texture features in the sub-window were extracted. Texture features were described by the Leung-Malik (LM) Filter Bank; (2) a strong classifier trained by AdaBoost algorithm was employed to assign every site a binary label. In order to train and evaluate the proposed method, a citrus dataset was provided which comprised 120 RGB images captured in natural illumination conditions. Twenty images in this dataset were used to train the AdaBoost classifier, and the remaining images were used as test set. Experiment on this test set showed that the proposed method achieved a precision of 0.867, and recall of 0.768, which confirms the validity of the proposed method.