Abstract:
The phenotypic information of soybean has been one of the most important indicators for the variety selection of soybean. Most research has been focused on soybean pods for phenotypic traits at present. However, the phenotypic information of soybean plants can also be a very important indicator for soybean seed breeding. Furthermore, the current manual measurement cannot fully meet the large-scale production in recent years, due to the time and labor-consuming. More recently, computer technologies have been started to automatically acquire the soybean phenotypic parameters. For instance, some shallow machine learning and image description were used for the image feature extraction and target detection, such as the Scale-Invariant Feature Transform (SIFT), Histogram of Oriented Gradient (HOG), and Support Vector Machine (SVM). Although the automatic acquisition of phenotypic data was realized to a certain extent, the high generalization ability can be a high demand suitable for practical application scenarios. In this study, an improved model was proposed to rapidly and accurately acquire the soybean plant phenotype using Re-YOLOv5 and area search algorithm. The circular smooth labels were also introduced in the Re-YOLOv5 to process the angle information for the detection of the rotating objects. The angle regression was then converted into a simple classification. As such, the rotating objects were accurately detected to reduce the redundant information in the detection area during object detection. In addition, the Coordinate attention mechanism was added to the Neck part of YOLOv5 (CSL). The cross-channel information was then obtained to capture the position and orientation information. As such, the improved network was used to more accurately locate and recognize the soybean branching than before. The weights greatly contributed to the YOLOv5 on the processing of feature-related parts. More importantly, the original 3×3 convolution kernels were placed in the backbone with a RepVGG block structure. The feature information was then fused to extract using different convolution modules. The information extraction of the overall structure was enhanced for the parallel fusion of the multiple convolution layers while reducing the number of model parameters. Taking the detected branch as the search area, an area search algorithm was also proposed to input into the algorithm, in order to extract the relevant information of the stem nodes in the area, and then connect the nodes in sequence. Thus, the soybean skeleton was reconstructed to obtain phenotypic information about soybean. The experimental results showed that the improved Re-YOLOv5 performed better to detect the rotating objects in various phenotypic indicators, compared with the YOLOv5. Specifically, the mAP of the improved Re-YOLOv5 increased by 1.70 percentage points, the number of parameters decreased by 0.17M, and the detection accuracy of stem nodes was improved by 9.90 percentage points. An excellent ability was also achieved to detect the small targets suitable for the acquisition of the soybean plant phenotype information. Among them, the average absolute errors of plant height, and the number of stem nodes and branches were 2.06 cm, 1.37, and 0.03, respectively, fully meeting the accuracy requirements of actual collection. At the same time, the detection area search algorithm can also be expected to accurately locate the stem nodes on each branch for the single-, double-, and complex-branched soybean plants, and then reconstruct an accurate soybean skeleton. Anyway, the improved model can also be used to accurately and efficiently acquire the phenotypic information of soybean branch angle, and soybean plant type. The finding can provide a strong reference for the subsequent acquisition of phenotypic information during soybean production.