Design of a winter-jujube grading robot based on machine vision

Abstract Winter-jujube is a kind of fresh fruit in China. After harvest, winter-jujubes require to grade into different categories according to their maturity levels. Mature winter-jujube can be recognized by their red colour. In this study, a winter-jujube grading robot is designed. Moreover, a method combining YOLOv3 algorithm and hand-engineered features is developed to calculate the maturity of winter-jujube. The grading robot is composed of a transmission unit, an image acquisition unit, and an actuator unit. Based on YOLOv3 algorithm, a detection model is trained, and compared with SSD and Faster R-CNN algorithms. When the IoU are 0.7, 0.8, and 0.9, the F1 scores of the model are 100%, 100%, and 93.66%, respectively. The mAP (IoU = 0.50:0.05:0.95) of the model is 94.78%, and the detection time of single image is 0.042 s. The detection model exhibits a high stability under different lighting conditions. In addition, overlapping winter-jujubes in the image can be detected accurately. After image distortion correction and object detection, an image processing flow for spatial positioning, size measurement and maturity calculation for winter-jujubes is designed. Finally, a real-time grading device for winter-jujube is built to perform grading experiments. The maturity grading accuracy is 97.28%, and the average grading time of each winter-jujube is 1.39 s.