Labour-saving detection of hybrid rice rows at the pollination stage based on a multi-perturbed semi-supervised model

Autonomous navigation of the hybrid rice pollination process is essential to increasing seed production. Detecting crop rows in real-time with flexible and cost-effective machine vision equipment is crucial to achieving autonomous navigation. The use of deep learning-based models has proven effective in crop row detection. However, training such models is highly dependent on manually finely annotated image data, which becomes a bottleneck that hinders their practical application and improvement of crop row detection accuracy. This study proposed a multi-perturbed semi-supervised learning-based model to enhance the semantic segmentation performance of hybrid rice regions by mining valuable information from easily accessible unlabelled images. The input data perturbation and network perturbation developed effectively alleviated the overfitting and teacher-student weight coupling issues when applying semi-supervised models to the hybrid rice image dataset with high content similarity. Navigation centrelines were obtained by post-processing the segmentation masks of the crop regions. Under a 1/2 partition protocol of labelled and unlabelled images, the semantic segmentation performance of the proposed approach achieved a mean intersection over union (mIoU) of 0.887, outperformed its supervised baseline (DeepLabv3+ ) and original model (mean teacher) by 1.8% and 4.1%, respectively. 82.38% of the crop row were accurately detected using the proposed model, demonstrating an improvement of 3.04% and 14.06% over its baseline and the original model, respectively.