High-Throughput Classification and Counting of Vegetable Soybean Pods Based on Deep Learning

Abstract

Accurate identification of soybean pods is an important prerequisite for obtaining phenotypic traits such as effective pod number and seed number per plant. However, traditional image-processing methods are sensitive to light intensity, and feature-extraction methods are complex and unstable, which are not suitable for pod multi-classification tasks. In the context of smart agriculture, many experts and scholars use deep learning algorithm methods to obtain the phenotype of soybean pods, but empty pods and aborted seeds are often ignored in pod classification, resulting in certain errors in counting results. Therefore, a new classification method based on the number of effective and abortive seeds in soybean pods is proposed in this paper, and the non-maximum suppression parameters are adjusted. Finally, the method is verified. The results show that our classification counting method can effectively reduce the errors in pod and seed counting. At the same time, this paper designs a pod dataset based on multi-device capture, in which the training dataset after data augmentation has a total of 3216 images, and the distortion image test dataset, the high-density pods image test dataset, and the low-pixel image test dataset include 90 images, respectively. Finally, four object-detection models, Faster R-CNN, YOLOv3, YOLOv4, and YOLOX, are trained on the training dataset, and the recognition performance on the three test datasets is compared to select the best model. Among them, YOLOX has the best comprehensive performance, with a mean average accuracy (mAP) of 98.24%, 91.80%, and 90.27%, respectively. Experimental results show that our algorithm can quickly and accurately achieve the high-throughput counting of pods and seeds, and improve the efficiency of indoor seed testing of soybeans.