Online sorting of drilled lotus seeds using deep learning

Abstract

For defective drilled lotus seeds, the inner bitter lotus plumule cannot be removed normally, leading to difficulties in subsequent nutrient extraction and food processing. There is an obvious difference in visibility of drilled hole between normal and defective drilled lotus seeds in the top view; thus, an online sorting method for drilled lotus seeds based on drilled hole detection is proposed in this study. First, a drilled hole detection model based on You Only Look Once (YOLOv3) is developed to detect the drilled hole features on the lotus seed surface. The model was tested and compared with the Faster Region-based Convolutional Neural Network (Faster R-CNN) and Single Shot MultiBox Detector (SSD) models, and it showed a better comprehensive performance in terms of accuracy and speed. A sorting control algorithm is also proposed to perform online sorting based on real-time drilled hole detection results. In addition, an auxiliary algorithm is proposed to prevent the boundary misjudgement of detected hole ownership between adjacent lotus seeds during continuous sorting. An online sorting system was designed, and sorting tests were performed. A sorting accuracy of 95.8% was achieved for the mixed defective and normal drilled lotus seed samples. The proposed method is expected to not only fulfil the practical requirements for the online sorting of drilled lotus seeds but also provide references for other agricultural products that require continuous online sorting based on the detection of local features.