Data segmentation method based on seedling density to improve the accuracy of rapeseed seedling recognition

The accurate recognition of rapeseed seedlings is the premise of achieving accurate counting. Traditional target detection algorithms and current improved algorithms cannot balance the counting accuracy and seedling recognition at different densities. Thus, this study proposed improved YOLOv7 recognition models by automatically dividing dense and non-dense datasets. For the non-dense dataset, the attention mechanism CA is introduced, and the loss function DIoU is substituted; for the dense dataset, the attention mechanism CBAM is introduced, and the loss function GIoU is substituted, so that the network focuses on the target object to improve the accuracy of the model. The results show that the improved accuracy of the non-dense area reaches 97.78%, mAP@0.5 reaches 98.32%, and mAP@0.5:0.95 reaches 66.86%. The accuracy of the improved dense area reaches 97.85%, mAP@0.5 reaches 95.91%, and mAP@0.5:0.95 reaches 69.9%. In addition, this paper compared SSD, YOLOv5, YOLOv7, YOLOv7-tiny, and the improved YOLOv7 and validates the feasibility of the improved YOLOv7 model. The improved YOLOv7 model proposed in this paper can provide a useful reference for future research directions of seedling recognition.