Citrus recognition in real scenarios based on machine vision

Abstract

At present, citrus is mainly harvested manually with low efficiency and high cost, thereby resulting in an imminent demand for fruit-harvesting robot. Hence, recognizing and locating citrus from complex backgrounds using machine vision technology is the premise and a key technology for robots to harvest them. To accurately and immediately recognize mature citrus, this study proposes a novel method based on the Otsu adaptive threshold method and on the improved random ring method. First, a citrus region was segmented using the Otsu adaptive threshold method overthe V component of the Luminance Chrominance (YUV) color space, extracting then the real citrus region after undergoing a denoising processing. Second, the citrus edge was extracted using a Canny detection operator with the curvature catastrophe point removed. The short citrus edge was placed in the background to obtain the citrus edge of each segment. Lastly, the citrus feature circle was fitted for continuous citrus edge in each segment using the improved random ring method based on the greedy algorithm. Experiments were conducted using the citrus images in various real scenarios. Results demonstrate that the proposed method increases the overall recognition rate of citrus to 95% with time consumption of approximately 80 ms. The proposed method improves the recognition efficiency, enhances the accuracy of citrus recognition, and provides technical support for fruit-harvesting robots. Lastly, the proposed method can be generalized to the location and recognizition of other similar fruits, thereby providing significant research value.