Immature green citrus detection by multiple features using machine vision

Abstract

Abstract. Forecasting the number of immature green citrus and its size prior to harvesting period would help the growers plan application of nutrients during the fruit maturing stages and estimate their yield and profit. Yield mapping by machine-aided harvesting requires automatic detection and the counting of fruit in a tree canopy. However, occlusion, varying illumination, and color similarity with the background make green citrus fruit identification a very challenging task. The overall goals of this research were to develop a real-time yield mapping system that could detect immature green citrus using color images with multiple features and to estimate yield in the complex outdoor environment. This study concentrated on estimating number of citrus fruit in images of tree canopy acquired under natural outdoor conditions. First, a linear color space, OHTA, was used to segment green citrus from the complex background. In order to overcome recognition challenges in uncontrolled environments caused by uneven illumination conditions, similar background features, and partial occlusion, multiple features such as color, shape and texture features of the different objects (immature green citrus, leaves, soil, sky, twigs and branches) were used for feature extraction to find differences between green citrus and background. Texture classification, based on two major types of texture features, gray gradient co-occurrence matrix and Tamura texture features, combined with a support vector machine (SVM) were used to remove non-citrus regions. After removing false positives, a final decision was made using shape features to detect individual citrus fruit. The proposed method was tested in a validation set of images obtained from the outdoor environment, detecting and counting over 82.4% of citrus fruit. The algorithm mentioned in this paper was promising in green citrus detection although improvement would be needed to decrease missed fruit and false positives.