Detecting endosperm cracks in soaked maize using μCT technology and R-YOLOv7-tiny

Abstract

Seed soaking is an important pre-sowing seed enhancement process in maize production. The efficient detection of invisible internal endosperm cracks during the soaking process is crucial for analysing the variations in maize endosperm cracks. In this study, a maize endosperm crack detection method is proposed using μCT technology and the R-YOLOv7-tiny model. The YOLOv7-tiny model was improved by introducing a rotated box, CSL, and Skew-NMS, transforming it into a rotated object detection algorithm. The GhostConv module replaced the Conv module, the SiLU activation function replaced the LeakyReLU activation function, and the CoT block and C3_TR module were added to the backbone and neck parts of the model to make the model focus more on crack feature extraction. Crack information was automatically extracted by integrating the crack length and number extraction algorithm in the detection head of the R-YOLOv7-tiny model. After 300 iterations of training, theP, R, AP, size, and detection speed of the model on a test set was 93.80%, 87.90%, 92.10%, 9.70 MB, and 67.11 fps, respectively. The proposed method was tested on endosperm crack images of six types of maize, and the MSE, RMSE, MAE, MAPE, and crack miss detection rate of the overall crack detection was 0.05 mm, 0.22 mm, 0.16 mm, 7.86%, and 7.29%, respectively. These results show that the proposed method could automatically extract information on endosperm cracks in seed-soaked maize with high accuracy. The results of this study can be utilised to analyse the changes in endosperm cracking during the soaking process accurately and rapidly, providing a reference for the maize soaking process.