Context. Rice (Oryza sativa) panicle provides important information to improve production efficiency, optimise resources, and aid in successful breeding of high-performing rice varieties.

Aims. In order to efficiently count rice panicles, a rice panicle recognition model based on YOLOv5s-Slim Neck-GhostNet was evaluated.

Methods. We used the developmental stages from heading to maturity as the time period to collect data for testing and validating the model. The GSConv convolution module from the YOLOv5 (You Only Look Once) model was compared with the original Conv convolution. We improved the original C3 module and replaced it with VoVGSCSP module, which further enhanced the detection ability of the model for small targets, such as rice panicles. To further optimise the performance of the model and reduce the computational complexity, we replaced the original backbone network of the model with a lightweight and efficient GhostNet structure.

Key results. Our results showed that the precision of the test set was 96.5%, the recall was 94.6%, the F1-score was 95.5%, and the mAP@0.5 was 97.2%. Compared with the original YOLOv5s model, mAP@0.5 increased by 1.8%, and the model size is reduced by 5.7M.

Conclusions. The improved YOLOv5 model had increased capability to detect and count rice panicles in real time. Our method reduced the size of the model while maintaining an acceptable level of accuracy.

Implications. The technology provides an intelligent and automated solution to better monitor rice panicle development, and has the potential for practical application in agricultural settings.